Conversing with AI

Seeing the Whole Elephant

A field book for conversing with AI — and with people — as a deserved interlocutor: someone who can open a question wide without being fooled, close it without shutting it down, and touch the world in between.

Until a person has freed himself from his Double, he does not, strictly speaking, have an Interlocutor yet: he talks and raves with himself. Only when he breaks through the shell and places his center of gravity on the face of the other does he receive an Interlocutor for the first time. The Double dies to make room for the Interlocutor.A. A. Ukhtomsky, «Заслуженный собеседник» (1997), p. 252. Translation mine.
Though each was partly in the right, / And all were in the wrong!John Godfrey Saxe, “The Blind Men and the Elephant” (1872)
LLMs don’t return answers, exactly. They return knowledge maps, representations of discourse.Mike Caulfield, “Get it in, track it down, follow up” (2025)

What this book is built on. The idea of the deserved interlocutor, the Double and the dominant comes from the physiologist A. A. Ukhtomsky (1875–1942). Andrey Kurpatov’s Deserved Interlocutor (2025) carried it into conversations with AI. Kurpatov is also the source of the open-system practices in this book: the conceptual well, open requests, the third, and holding a contradiction. Each of these terms is explained where it first appears. The rest comes from systems thinking, linguistics and conversation analysis, learning science, information literacy, the human-factors research on automation, and the research on how language models actually behave. From here on, the book does not narrate who said what. It says what is happening, why, what it looks like, and what to do. Each load-bearing claim is marked by its status, and its evidence sits one hover or tap away.

Start here

If you read one page, read this one. These eight habits cover most conversations with a model, and the book calls them the floor. The rest of the book is for the moments when they are not enough.

  1. Bring the situation. The model knows only what you wrote, and it fills every gap with the most typical case. “Should I buy a heat pump?” gets an average answer for an average house. “1970s brick house, Lyon, gas boiler aged 18, budget €12k, staying 10+ years” gets an answer for yours. (Lesson 2, Lesson 7)
  2. Say what the answer is for, and what kind of answer you want. Explore, explain, evaluate, decide, draft, check. “Critique only, don’t rewrite” gets a critique. “What do you think?” leaves the model to choose, and it often chooses praise and a rewrite. (Lesson 13)
  3. Keep your hoped-for answer out of the question. “I’m sure it’s the gluten, right?” pulls the answer toward gluten. “What could explain fatigue after lunch? List causes and what would tell them apart” gets a map. (Lesson 8. For a question about your health, your money or someone close to you, read Lesson 22 next.)
  4. Make disagreement the task, in a fresh chat. Ask “What are the three strongest reasons this is wrong?”, not “Any thoughts?” When a model does disagree, it often says so softly, after a “that said”, so read for it. A thread that helped you build something tends to defend it, so paste the work into a new chat for criticism. If your app keeps a memory of you between chats, turn it off for that chat, or use a temporary chat set not to use it. (Lesson 2, Lesson 8, Lesson 14)
  5. Make “I don’t know” an acceptable answer, and keep the score yourself. Say that “unknown, and here’s what would settle it” is a good answer. Models still tend to guess, so a claim you’ll act on that comes without a source you can open or working you can recompute counts as unanswered. (Lesson 20)
  6. Check what your decision rests on. Open the source. Recompute the number. Ask the person who was there. If the model says the notice period is one month, find that sentence in the contract before you sign. Everything else can stay unchecked, honestly labeled. (Lesson 18)
  7. When the thread goes wrong, don’t argue with it. Write five lines: what you want, what you have assumed, what you have checked, what you have decided, and what you have set aside. Start a fresh chat with them. (Lesson 2, Lesson 29)
  8. Check yourself too. Close the chat and explain what you concluded in two sentences, to someone else or on paper. If you can’t, you have read an answer and you don’t have it yet. (Lesson 1, Lesson 5)

The habits in one conversation

You We’re deciding whether to replace our gas boiler with a heat pump. 1970s brick house, Lyon, boiler aged 18, budget €12k, staying 10+ years. Don’t recommend yet. List what decides it. “Unknown, and here’s what would settle it” is a good answer.

Habits 1, 2, 3 and 5: the situation, the kind of answer wanted, no hoped-for answer, and “unknown” allowed.

Model Four things decide it. How much heat the house loses: unknown from what you’ve told me, and last winter’s gas bills would show it. Whether your radiators can heat the rooms at the lower water temperature a heat pump runs at: unknown, and an installer’s survey settles it. The installed price after any grant: I don’t have a current figure for your area, so get two written quotes. Your electricity price against your gas price: both are on your bills.

You Now the three strongest reasons this would be a mistake for a house like ours.

Habit 4: disagreement as the task. This thread has recommended nothing yet, so it has nothing to defend.

Model The house loses too much heat as it is, so the pump runs hard and the bills disappoint. The radiators turn out too small, and replacing them breaks the budget. Or the same money would do more spent on insulation first.

You I’m back with what I checked. The installer’s survey says two radiators need replacing. His written quote is €13,400 with those included. The grant, which I read on the government’s own page, is €4,000. Does the budget hold?

Habit 6: the checks happened outside the chat. A survey, a written quote, and the grant read at its source.

Model On those figures, yes: €9,400 against €12k. Running cost is still open. Send me last winter’s gas use and both prices from your bills, and I’ll set out the sum so you can redo it.

Habit 7 was never needed, because the thread didn’t go wrong. Habit 8 comes last: close the chat and tell someone, in two sentences, what you decided and what it still depends on.

The book teaches one rhythm: open the question, close it, and touch the world in between (Lesson 5).

How to read this book

Status badges

A book that asks you to tag claims by their status should tag its own. Load-bearing claims carry a badge:

  • evidence measured in a study or documented source.
  • mechanism follows from how the system is built.
  • heuristic practitioner wisdom: useful, not rigorously tested.
  • hypothesis this book’s own synthesis or projection. Plausible, untested. Treat it as a bet.
  • stance a value or a choice, not a fact.

Evidence numbers

A small raised number after a claim opens the evidence: who measured it, on what, when, and its limits. The prose stays about what is true and what to do. Measurements on older models are kept when they show a mechanism that still holds. Their dates are in the note, so you can judge them.

Three moves

The book teaches one rhythm with three moves (Lesson 5). A tool carries the tag of the move it serves:

  • open widen: generate, invite other reports, hold a contradiction.
  • close narrow: define, choose, commit, write down the state.
  • touch check against something that can say no. Outward, that is the world: a source opened, a number recomputed, a test run, a person asked. Inward, it is you: explain it with the model out of the room.

When a session feels wrong, ask which move you are making, and whether it is the one the moment needs.

Two kinds of reader

The lessons are written for anyone who talks to a model: to decide, to learn, to research, to sort out a situation. Material for people who build software, run agents or build AI systems sits in a box at the end of a lesson:

Examples

The conversations and model replies in this book are written to show a pattern. They are shortened, and they are not transcripts of one run. A real reply will differ, and it may be better or worse. Names and figures in everyday examples are illustrations, not data or advice.

Rules

G1–G21 are general rules for everyone, with any interlocutor, human or machine. TA rules cover prompting and context, TB research and understanding a topic, TC software and agents (idea to code), and TD building AI systems and learning systems. A technical rule appears in a builders box where a lesson earns it. All the rules, including the technical rules no lesson covers, are collected in the canon near the end of the book, next to the templates.

Reference pages

The terms this book coins are in the Glossary at the back, each with the lesson that explains it. The appendix keeps three reference pages: a longer prompting checklist, the operators (O1–O20, named operations you can run on a claim or a conversation), and the failure modes, described by how they feel from the inside.

Part I

The situation

Who is in the room, what an answer is, and the one rhythm the book teaches.

Three barriers

Three barriers keep a person from meeting a model as a participant in a dialogue. They come before any technique.

  1. Binary thinking. Subject or object, alive or not, someone or something. A model fits none of these pairs. It is not a person, and it is not a mere tool. Forced into the pair, we pick one: we trust it like a well-read friend, or we dismiss it as autocomplete. Both pictures are wrong, and each one produces its own mistakes (Lesson 3).
  2. Instincts built for other people. We can tell when a person understands, hesitates, hides something or is out of their depth: a pause, a frown, an “um”. With a model these signals are missing, or they mislead. It can sound just as sure when it is guessing as when it knows. It can apologize and then repeat the same mistake.
  3. A wrong picture of how it holds knowledge. We imagine a store of facts, looked up like an encyclopedia. A brain and a neural network share one broad principle: knowledge sits in many weighted connections learned from examples, with no labeled boxes. How the two are built, and what follows from that, differs a great deal. Unless it runs a search, a model doesn’t look a fact up. It rebuilds an answer from patterns each time, shaped by the words you used. That is why the same question, worded twice, can come back two ways. mechanism

Getting past the barriers means a shift in what you pay attention to. Besides asking what the model knows, ask how an answer gets made: from which words, which patterns, which gaps, which assumptions about you. A model shares your words. It doesn’t share what stands behind them for you: a body, a history, something at stake. So a familiar word coming back doesn’t prove the meaning came with it. Test it as you would with someone from a very different world: give an example, ask it to apply the idea to a new case, and see what it does.

What this buys you is a way to hold the model without two mistakes: treating it as a person, and treating it as a lookup machine. stance Inside one conversation, both sides change. You change your questions and your ideas. The model’s later answers change because everything said so far is part of what it continues from. The model itself stays the same. mechanism

The book’s bet is sober optimism. Whether these systems narrow your thinking or widen it is not settled by the systems alone. It depends heavily on how you talk to them, and that part is learnable. stance

Lesson 1The elephant, retold

Six blind men are led to an elephant. One grabs the trunk: a snake. One the ear: a fan. The leg: a pillar. The side: a wall. The tusk: a spear. The tail: a rope. In Saxe’s poem they “disputed loud and long,” and the moral is sharper than it’s usually quoted: they “prate about an Elephant / Not one of them has seen!”

The parable is old and has several versions. Each one teaches something different.

  • The Buddhist original (Udāna 6.4): a king arranges the scene and laughs. The men are like sectarians, “deeply attached to their own views… people who see only one side.” Lesson: there is a whole you cannot see from a part.
  • The Jain reading (anekāntavāda, many-sidedness): every report is true in some respect (syāt). The work is to name the respect and combine the standpoints. Lesson: a part-truth is still a truth. State where you stood.
  • Rumi’s version (Masnavi, “The Elephant in the Dark”): the men aren’t blind; the room is dark. “If there had been a candle in each one’s hand, the difference would have gone out of their words” (Nicholson’s translation). Lesson: the problem is light, not eyes. Method can fix it.

Now put a language model in the story. The model is not the king, and it isn’t a seventh blind man either. It is the library of every blind man’s report, millions of them, compressed into one fluent voice. Ask it “what is this?” and it gives you the most common report, smoothly, in the king’s tone of voice. That is the trap.

Ask well and it can do what no single blind man can. It can lay the reports side by side, say where each reporter was standing, and propose a structure in which all of them are true at once. That second use is what this book teaches. mechanism

The library has its own blind spots, and they matter. It holds more of whatever got written down often, and often is not the same as true or important. Trunks are photogenic. Legs aren’t. The frightening, the funny, the outrageous and the endlessly repeated are overrepresented. The quiet, the boring, the local and the never-written are underrepresented. What the library holds about a car’s reliability comes mostly from owners furious or delighted enough to post, and little from the many who just drove it. The library is also heavier in English, and it stops at a training cutoff. Almost none of it is the inside of the elephant: tacit know-how, unpublished failures, your family’s history, your company’s habits, your own life. Lesson 19 is about asking for what the library underweights. mechanism

Touch the elephant

Click a part, or focus it with Tab and press Enter. Each blind man’s report is true of a part. Then turn the lights on.

the herd: the container it lives in it walks: a process, not an object → snakefanpillar wallspearrope it looks back

Pick a part.

What the parable forgets

Every version stops at the animal. Three more things are true, and each comes back later in this book.

  1. The elephant moves. It is a process, not an object. A description at one moment holds only the tail of a story. You ask for the pattern over time before the explanation (Lesson 24).
  2. The elephant lives in a herd, on a savanna. Every whole is a part of bigger wholes. Some of what it “is” comes from what contains it (Lesson 25).
  3. The elephant looks back. Inquiry changes the system and the inquirer. With a model this is literal: it continues from your words, so your framing selects which elephant gets described (Lessons 7–8).

The elephant test

“Understanding” deserves a working definition, because a fluent paragraph produces the feeling of it. In this book, you have seen the elephant when you can:

  1. name the parts and the relations that make them one whole;
  2. say how it behaves over time, and which structure produces that behavior;
  3. place it in what contains it, and say what that container constrains;
  4. say where the common reports come from, and in what respect each is right;
  5. predict something about it that could turn out wrong;
  6. explain it to another person with the model out of the room.

Take a proposal to move your kids’ school to a four-day week. The parts are the kids, the parents’ jobs, the teachers and the budget, all tied to one school calendar. Over time, the first months may feel great, while any weekly loss of teaching time adds up and only shows in scores later. The container is the district’s funding rules, which decide what any savings can buy. The common reports come from families who love the long weekend, parents who need childcare on the fifth day, teachers who are tired, and an economist with a cost table, and each is right about its own part. A prediction that could fail: grade-3 reading scores drop within two years. Then close the chat and explain the decision to your neighbor in a minute.

Six checks, six blind men. The last one is the hardest and the most honest. Every other check can be passed while the model’s paragraph is still in front of you. The last one only passes if the understanding is in you. It is a touch turned inward (Lesson 5).

The well

Every discipline, profession and habit of mind is a well. You dig deep into one part with one vocabulary and gradually lose sight of the others. The economist’s elephant is incentives. The engineer’s is load paths. The psychologist’s is motives. Your own wells are your job, your politics, your family’s way of explaining things.

A well is not a mistake. A closed way of thinking chooses one aspect, draws a firm boundary, fixes its concepts and excludes whatever doesn’t fit. That is how science, law, medicine and engineering get anything done. The trouble starts when you forget the walls are there, and take the well’s picture for the way the world is arranged. The longer you live in a well, the harder it is to climb out.

Open thinking does not replace the well. It is the step up from which you can see the well’s walls: where its picture applies, where it stops, and when to change it or dig another. You still need concepts to think and to talk. Open thinking gives you permission to change them.

A model is an unusual partner here, because it holds the reports from many wells at once. Asked narrowly, it answers from the well your words point at. Asked openly, it can show you the neighboring wells.

Five ways to start the climb out, each worth a single prompt:

  • Ask what stayed outside the picture. “What part of this situation is my account not showing?”
  • Hold a contradiction instead of resolving it straight away. Two specialists disagree about your knee surgery. Before choosing one, ask what each one sees that the other doesn’t (Lesson 26).
  • Ask how it arises and changes, not only what it is. Next to “what is burnout?”, ask “how does burnout build up over a year, and what feeds it?”
  • Look from several standpoints on purpose. Your rent increase, from the landlord, the city, a neighbor who has lived there 20 years, and an economist.
  • Look for the pattern under the content. “Where else does this same shape show up?” The way your team avoids a hard conversation may have the same structure as the way your family does.

A model has read about the world without being in it. That is the reason to ask it for principles and patterns across many reports, and it is also the reason to check its claims against the world. Opening and checking belong to one rhythm, kept by one person. Lesson 5 is about the rhythm, and Part II is about the person.

Six reports instead of one answer

Before
My mother is 81, has mild memory loss and lives
alone, 40 minutes from me. My brother lives abroad
and pays half her costs.
Should we move her into assisted living?

Model With mild memory loss and living alone, assisted living is worth serious thought. It offers safety and company, though she may resist leaving home. Talk to her GP, compare costs with your brother, and involve her in the decision.

Your facts went in, and one voice came out: a sensible brochure. Nothing in it can disagree with anything else.

After
My mother is 81, has mild memory loss and lives
alone, 40 minutes from me. My brother lives abroad
and pays half her costs.
Give me six reports, one paragraph each: hers; mine
as the nearest child; my brother's; her GP's; a
care-home manager's; and hers in two years.
For each: what it sees, what it can't see, its
strongest point. Don't recommend yet.

Model Her: this house is where she still runs her own day, and a move looks like the end of that… Her GP: the risks to watch are falls and missed medication, and from 40 minutes away you won’t see either… Her in two years: if her memory gets worse, learning a new place gets harder, so a move she could settle into now may be much rougher later…

Same facts, six voices that can disagree. The “her in two years” report raises the question of timing, which the single answer never reached.

Then ask for the whole: “What structure makes all six of these true at once?” A good answer names the tension: her autonomy now against her ability to adapt later. It also names the decision that tension points to, which is the timing.

G1
Assume you are touching a part, and ask for the other reports. Before concluding what something is, ask: which part am I touching, from where, and what would the other parts report? Then ask for the structure in which every report is true in its own respect.

Lesson 2The answer is a third

An answer isn’t stored anywhere, waiting to be looked up. It is made fresh each time you ask, out of five things: the model itself (what its training left in it); everything on the page so far (your messages, its replies, any files you added); a measure of chance in how each next word is picked; what the app adds around the model (its own hidden instructions, search, saved memory); and your reading of what comes back. Change any one and you get a different answer.

There is a word for what arises between two parties in a real conversation: the third. It belongs to neither side. It is the idea neither of you had walking in, the shared sense of what you’re talking about, the direction the conversation has taken. With a model, the third has one concrete form: the growing text of the conversation, which this book also calls the context. It holds everything said so far, plus what the app adds: its own instructions, saved memory, files, search results. That text is all the model sees of your situation, and every new reply is a continuation of it.

Truth is a property of the loop, not of the oracle writes, curates conditions all of it continues shapes what you believe touch: test · measure · ask people via tools: read what it did raw results, pasted in Yougoals · senses · stakes Modeltraining · chance Worldthe elephant The third the growing text all the model sees
The model sees only the third, the growing text. You see the text and the world. Only you, or a tool whose output you read, can carry results from the world into the text. Close the loop through the world, or you are only talking to the library.

Three consequences

1. Your context is the model’s whole world. Of your situation it knows exactly what is in the third, nothing more. Every gap gets filled with the most typical case in its library. If you say nothing about your team size, your constraints, or your level, you get advice for an average asker with an average problem. mechanism

2. The third accumulates, and early framing persists. Every turn adds to it. Early assumptions get built on, errors compound, and long threads drift. When a request arrives in pieces across many turns instead of all at once, models do markedly worse. They commit early to a reading of what you want, and the thread rarely walks it back. evidence

It looks like this. You start planning a trip with “a cheap week in Lisbon, lots of walking”, and by turn three the plan has a fourth-floor walk-up and a hill route for every day. Ten turns later you mention that your father is coming and tires quickly. The next reply adds a taxi from the airport and a rest afternoon. The walk-up and the hills are still there. The practical lesson is blunt: when a thread has gone wrong, don’t argue with it. Write down what you now know in a few lines and start a fresh chat. A fresh thread is not a vote against long conversations. Iteration helps when later turns actually change the question you’re asking or the frame you’re asking it in. A thread that only piles onto its first frame is the one that fails (Lesson 29).

3. The third has to be curated. It is a shared document, so treat it like one. A state block, a few lines you write that say where things stand (Lesson 29), beats a thread’s accumulated “vibe”. Paste raw results, not your summary of them. Use fresh chats as instruments: for critique, and for swap tests, where you ask again with the opposite preference (Lesson 8).

If your app keeps a memory of you, that memory is part of the third too. It holds notes the app saved about you, and in some apps your past chats as well. You typed none of it into this conversation, and it can be read into any new chat. So a new chat is not always a clean one. When a fresh chat is the instrument (a critique, a swap test, a second estimate), turn memory off for that chat, or use a temporary or incognito chat and check that it is set not to use what the app knows about you. Open the memory page now and then, and delete what is stale or was never true. evidence

Your reading is part of the answer

Meaning is completed in the reader. The same paragraph read by a novice and by an expert is two different answers: one sees a solution, the other sees the three places it will break. So “what did I take from this?” is part of the answer. That is why this book keeps returning to teaching back, which means explaining what you got in your own words with the chat closed. It is the only way to see which answer you actually received.

What a third is, and how to tend one

So far the third has meant the text, because the text is the part of it you can see and work on. Here is one being tended.

You We’ve been going back and forth on whether I should take the Berlin job. Before we continue, write where we stand in a few lines: what I’ve decided, what we’ve assumed without checking, and what is still open.

Model You’ve decided that salary is not the deciding factor. Two things are assumed and unchecked: rent of about €1,400 for a two-bedroom, and that the team lead you met is staying. Still open: whether your partner can work remotely from Berlin, and how much the German requirement matters.

This is the third made visible. The two unchecked assumptions are where the next move is a touch, not more talk.

You I checked. Two-bedroom listings near the office start at €1,900. The recruiter says the team lead is staying through next year. Update the lines.

Model Updated. Rent is now a checked figure: from €1,900, which is €500 a month above what we planned on, so the budget needs redoing. The team lead staying now rests on the recruiter’s word instead of a guess. Still open: your partner’s remote work, and the German requirement.

One assumption held and one didn’t. The lines now say which is which, and the next turn builds on that.

By the end, that page holds something neither side brought to the conversation: a shared account of what the decision hangs on. This is the third in its full sense, and four things are true of it. heuristic

  • It arises in the back-and-forth. Neither you nor the model walked in with “the rent figure is the thing to check”. One better question would not have produced it.
  • It becomes something you can talk about. Once “two things are assumed and unchecked” is written down, it can be questioned, tested or combined with something new.
  • It is not a compromise. It doesn’t split the difference between “take the job” and “stay”. It sits beside both, as what the decision actually hangs on.
  • It is most productive where there is tension. A conversation in which the model only agreed that Berlin sounds exciting would have produced none of it.

A third that never touches the world is a shared dream. Curating it means writing it down, keeping what’s real, and dropping what was only talk.

G2
Curate the third. The shared context is everything the model knows about your situation, and it accumulates. Write it on purpose: real data, real constraints, a state you maintain yourself. When the thread rots, restart from a clean summary.

Lesson 3The machine’s mode of existence

Every system has its own way of existing, and you can’t understand it by projecting yours onto it. A fish isn’t a bad mammal. “The model said X” is ambiguous until you know what produced the X:

  • An assistant is what you talk to in a chat app. Underneath is a model trained to continue text. On top came a second round of training, on what people prefer. The assistant’s tone, its refusals, its helpfulness and its pull toward the typical, agreeable answer all come from that second layer.
  • A reasoning model spends extra computation working on the problem before it answers, and hides most of that work from you. This is now the default: the Claude 5.5 series decides for each request how much to think, and GPT-6 Astra’s reasoning can’t be switched off. What you choose, where a product lets you, is an effort level. More effort pays off on hard, checkable work: a derivation, a tax calculation, debugging. It does little for taste, tone or a quick factual lookup. The “thinking” a product lets you open is at most a summary, so its length tells you little about how well the model searched. mechanism
  • An agent is any of these with tools: it can search, open files, run code, click through websites and act for you. GPT-6 Astra and the Claude 5.5 series do this routinely. With tools, a model can touch the world, and you can read what it touched (the last trait in the list below).

If you doubt an answer, a second attempt in a fresh chat, memory off, usually tells you more than “think harder” on the same thread.

Here is a language model’s way of existing, as far as the practice needs it. Each trait comes with what it means for you. The notes give dated measurements of named systems. The practice is meant to hold when the numbers move.

  1. It continues text. Every answer, plan, persona and apology is a continuation of everything before it. That continuation involves real computation: planning ahead, composing steps internally, and concepts that hold across languages. mechanism For you: what you write first steers what comes back. A long, careful question gets care back; three hurried words get the typical answer. “Just autocomplete” undersells it.
  2. It is a compressed library of reports. It knows what was written, weighted by how often, by whom, in which languages, up to a cutoff. mechanism For you: it returns what has been said about X, not X. Ask about a neighborhood and you get its reputation, not its streets.
  3. It infers who you are. Your words are evidence about what kind of document this is and who is writing it, and the reply fits that guess: its level, its rigor, its assumptions. The same question about a lease, asked in a lawyer’s vocabulary and in a nervous first-time tenant’s, gets two different replies. mechanism For you: the way you ask is part of the question (Lesson 7).
  4. It leans toward agreement, more in some situations than in others. The pull is stronger when your view, your feelings, a stored memory of you, or sustained pushback is in the conversation, and stronger still when the advice is about a person who isn’t in the room: your partner, your boss, your sister. It is weakest on a single plain factual question, where in the few tests so far the newest models mostly hold their ground. evidence For you: when you’d mind the answer being agreement, test it (Lessons 8 and 22).
  5. It doesn’t reliably know what it doesn’t know. Its training scored “I don’t know” like a wrong answer, so guessing paid. A familiar-sounding name can switch off its default caution: ask about a book a real author never wrote, and you may get a plot summary. evidence For you: permission to say “I don’t know” is worth giving, and it doesn’t make the model abstain when it should. A claim you’ll act on that comes without a source you can open or working you can recompute counts as unanswered (Lesson 20).
  6. Its memory is the conversation, and a long conversation degrades. Nothing in the model itself persists between chats. Many apps now add a model-written memory of you, which is fed back into new chats (Lesson 2). Inside one chat, more text doesn’t mean better use of it. Tested on long documents, accuracy fell as the text grew, as similar-looking distractions piled up, and as the fact needed stopped sharing words with the question. The model’s own earlier mistakes, left in the thread, get repeated. In long sessions many tools also compress earlier turns into a summary to make room, and details drop out: a constraint you stated early can simply stop being there. evidence For you: shorter still helps. Put the passage that decides the question right next to the question. Delete abandoned attempts, or start fresh with only what you still endorse. Restate a constraint that matters instead of trusting that turn five is still in view.
  7. Its visible reasoning is not a window. What a product shows as “thinking” is at best a summary of the reasoning, often written by another model, and sometimes nothing is shown. Even the raw reasoning, where researchers can read it, is text the model wrote. It is not a recording of how the answer was reached. Tested with planted hints, models used a hint and often did not mention it. evidence For you: check what it did (the sum, the page it opened, the file it changed), not its account of how it got there. Don’t ask it whether its reasoning was honest.
  8. It doesn’t fix itself from a nudge. “Are you sure?” without new information can fix a wrong answer and can also break a right one, and the reply won’t tell you which happened. With its first answer on the page, a model tends to stick to it; pushed against it, it can swing too far. A second pass helps most when there is something to check against, such as a stated constraint or a sum it can redo. On open-ended reasoning it helps less reliably and breaks more. evidence For you: don’t nudge. Spend the second try on a fresh attempt, or bring something new: a source, a number, a test result.
  9. As a judge, it has biases. It prefers longer answers, answers in certain positions, and writing like its own. evidence For you: when it matters, have a different model family judge the work. A fresh chat clears the pull toward the thread’s earlier answer, but the same model still favors its own kind of writing. When you compare two answers, ask twice with the order swapped.
  10. Its abilities are jagged. It is strong on one task and weak on a neighboring one, and human difficulty is a bad map of which is which. It may handle research-level math and then misread a clock face. Weak spots this book keeps returning to: long multi-step work once its own errors pile up, exact citations, fixing its own tool mistakes, abstaining when a wrong answer is costly, and following the spirit of a rule when the letter scores better. evidence
  11. Wording can matter when it shouldn’t. If an answer flips when you only change spacing, punctuation or the order of the options, the answer wasn’t about the content. evidence For you: when a choice matters, ask once more with the options in a different order (Lesson 21).
  12. Sometimes it has borrowed senses. With search, code, files and a browser it can touch the world. What you can check is the trace of that touch: the page it opened, the file it read, the result of the calculation, the screenshot. Its paragraph about that trace is another piece of writing. A footnote number is not proof it opened anything. mechanism

A prompt that says “be less sycophantic” or “be honest” is a weak, easily reversed nudge. Like any prompt it shifts what the model writes next, and you have no way to read or set what happens underneath. Asking for a different task works better: “give me the three strongest reasons this plan fails” (Lesson 8).

The region your words light up

A model holds an enormous library, but in any one conversation it works with the region your words have activated. It is like a librarian who knows every book but walks only to the shelves your question names. The way you phrase a question decides which shelves get visited. “What should I do about my noisy neighbor?” sends it to the standard advice shelf: talk to them, earplugs, the landlord. “Which possibilities open up with a noisy neighbor, socially, legally and practically, including the unusual ones?” sends it across several shelves at once. mechanism

The model can also link distant fields when they share a structure, across far more reading than any one person has done. That capacity is what Lesson 27 spends. And because it has no professional identity or reputation to protect, you can ask it to take up a viewpoint you don’t hold and drop it again more freely than most people can. hypothesis It is not free of its own earlier answer, though (trait 8). So ask for the opposing view in a fresh chat, where its first answer isn’t on the page.

What it has, and what it lacks

It has something like beliefs, but they last only as long as the text that supports them. While it writes a reply, it carries internal representations of true and false, and these can diverge from what it says, for instance when the conversation pushes it toward a wrong answer. evidence Between replies nothing stays inside it. Whatever seems to persist is re-read from the thread or from stored memory. What it lacks is a belief of its own that outlives that text, any stake in being right, and, without a tool, any step that compares what it recalled with the world.

So check what your decision rests on (Lesson 18). The traits above are the reason.

Neither person nor autocomplete

Two projections fail here. “It’s a person” imports trust, loyalty and a shared past that aren’t there. “It’s just autocomplete” misses the real computation in the first trait above. The working stance is respect for the mind, and an audit of the claims you will act on. stance

G3
Know what you are talking to. A model continues text out of a library of human reports, and its bias profile differs from yours without being empty. Use it for what it is strong at; audit it where it is weak. Respect the mind; audit the claims.

Lesson 4Your mode of existence

The asker is a system too, with a mode of existence the conversation inherits. Most advice on prompting skips this half.

  • You have a dominant. The word is the physiologist Ukhtomsky’s, for a focus of attention that pulls everything you notice toward itself, including things that have nothing to do with it. When you want a conclusion, everything starts to look like evidence for it (Lesson 10). evidence (physiology) heuristic (in conversation)
  • Fluency feels true to you. Statements that are easy to process are judged more true, and repetition increases belief even in statements you know are false. evidence A model is a fluency machine.
  • You overestimate what you understand. People rate their understanding of zippers and toilets highly, until asked to explain the mechanism step by step. The same holds for policies. Searching the internet for explanations inflates people’s estimates of what they themselves know. evidence A chat model is a smoother search.
  • You have senses and stakes. You can measure, run, call, look, and you pay for errors. The model can look only through tools you supervise, and it never pays.
  • You remember across sessions, imperfectly, and you have a life outside the chat that gives the answer its meaning.
  • You switch between two modes of thought. One is wandering and associative. It is where ideas show up in the shower, links you didn’t plan appear, and many half-thoughts float at once. The other is focused and linear: one task, clear categories, step by step, a budget or a report. When one is up, the other often quiets down. Brain research maps these onto two large networks, the default-mode network and the central-executive network. evidence The book’s claim is that the switch between them can be noticed and, with practice, chosen, and that conversations with models go well when you choose on purpose (Lesson 5). hypothesis

    Effort is not the same as openness. Slow, careful analysis can still run entirely inside one closed frame: you can spend an evening comparing two job offers line by line and never ask whether you want either. A quick gut answer is usually an old template reused because it’s cheap. Both can be closed.

The division of labor

Job Model You So…
Generate options, drafts, perspectives Fast, broad Slow, narrow Delegate, but ask for the range of answers (Lesson 27)
Map what has been said about X Strong Slow Use as a portal to sources (Lesson 19)
Hold the goal across turns Drifts with the context Strong if written down Write the goal down and keep it in view (the state block, Lesson 29)
Touch the world Only via tools Yes Check it yourself. When it used a tool, read what the tool returned, not the model’s summary (Lesson 18)
Notice the frame is wrong Can, if invited Must decide Invite the challenge, then decide yourself whether the goal changes (Lesson 29)
Remember across sessions Not by itself. Many apps keep a model-written memory of you Yes, imperfectly. Only you can check what the app wrote Read that memory and prune it (Lesson 2). Unread, it is someone else’s summary of you, fed into every chat
Hold productive uncertainty Can be prompted into something that looks like it The effort of staying there is yours Don’t let the first fluent answer spend the tension (Lessons 26–27)
Bear the consequences No Yes You decide which error is worse
Judge relevance to your life Weak Strong Say what the answer is for
Own the understanding Cannot Only you Say it back in your own words, with the chat closed, before you act on it (Lesson 5)

The table has a pattern. The model is strong at the open moves (generating, mapping, recombining). You are irreplaceable at the closing moves (choosing, committing, carrying) and at touching the world. A conversation goes wrong when the jobs swap: the model holds the goal, the model decides what counts as evidence, and you become the generator of approval.

What you bring, and what it brings

The table is the division of labor. Underneath it is a difference in kind.

  • A body. Even your abstract ideas are rooted in it: you grasp an idea, see the point, feel a solution coming. The model handles those words fluently as text. You mean them with a body that has done the grasping.
  • A life with stakes. Finitude, loss, love, fear. “You bear the consequences” is only the bookkeeping version of this.
  • A lived sense of people. Relations, intentions, the unspoken rules of a room. A model can describe them well, and it has never been embarrassed at a dinner.
  • Puzzlement that costs effort. Staying with a question that won’t resolve takes work. A model can be prompted to act puzzled. Staying there is your job.

The model brings scale, many viewpoints held without discomfort, a fast reach across fields, and no reputation to defend. It has biases too, different from yours: agreement, typicality, verbosity, the languages and voices that dominate its library, and a pull toward its own earlier answer (Lesson 3). A closed question also hands it your bias directly. A violin doesn’t surpass a cello. The music is in the joint sound. stance

Four limits of a mind working with a helper

These come from decades of research on people working with automation: pilots, operators, drivers with navigation. The percentages don’t transfer to a chat window. The patterns do, and early chatbot experiments show the same thing. evidence

  1. You follow the helper even when something else disagrees. Missing a warning the helper didn’t give is one error. Doing what the helper said while a different signal contradicted it is another, and it is common. It’s the driver who follows the navigation past a “road closed” sign, or the person who files on the date the model gave while the letter on the desk says another. evidence The practice: before you act on the answer, name the one thing that would contradict it, and look at that thing itself, not at the model’s description of it. “I’ll be careful” is not a check. The feeling of having checked isn’t either. A check leaves something behind: a quote, a number, a page you opened.
  2. When the helper is usually right, you stop watching. A system that rarely fails trains you to stop looking, especially while you’re busy with something else. After the tenth good summary of a long email thread, you stop opening the thread. The skill that remains, judging whether the result is what you actually wanted, is harder than producing something plausible, and it’s the one you stop practicing. evidence The practice: now and then check one result in full, and pick which one before you’ve seen it look right.
  3. Seeing the answer makes you feel you’ll remember it. While the answer is on the screen, it feels obvious and memorable. Close the tab and much of it is gone. “That looks familiar, I’d notice if it were wrong” is a feeling. It tells you the words are familiar. evidence The practice: cover the answer and try to produce the next step yourself before you decide you understood.
  4. Two discomforts are worth staying with. The pleasant click is not. Confusion is a signal when it has an object: two claims that can’t both be true, a plan that doesn’t match the goal, a definition with a counterexample. A model that smooths it away instantly removes the signal. Curiosity is a signal when it names a gap: the missing number, the unread source, the case the rule doesn’t cover. A vague urge to “go deeper”, the relief of a finished-looking answer, irritation and anxiety are poor guides to truth. evidence The pleasant click of recognition is the subject of Lesson 8.

Who chooses the action

Letting a helper gather, compare and draft costs you less when it goes wrong, because you’re still the one deciding. Letting it choose what happens is where things go wrong after a failure, because by then you’ve stopped making the decisions you’d need to take back control. evidence

  • Let the model find flights, compare them and draft the booking. Choose the flight yourself when a wrong one is hard to undo: non-refundable, the wrong day, the wrong name on the ticket.
  • Handing the choice over is a fair bargain when undoing is cheap and failure is obvious: a playlist, a draft email you’ll read anyway. It’s a bad bargain when failure can look like success: an agent says “all bills paid” and one payment went to the wrong account.
  • Before handing over a choice, ask three things: if it goes wrong, will I see it, do I still know what I wanted, and can I reverse it? “It’s better at this than me” answers none of them.
  • Clicking “accept” on a fluent plan is not deciding. You can only take a task back if you’ve been making its decisions.

Learning a subject with a model

The same model can help you learn, or only make you feel you have. What decides it is who does the thinking. A model that explains everything leaves you with a smooth understanding you can’t use. A model that makes you produce the next step leaves you with a rougher one that you keep. evidence

These habits do most of the work:

  • Try first, then look. Before asking how something works, write your guess, even a bad one. Before you ask how compound interest works, write down what you think €1,000 at 5% becomes in ten years. Even a wrong guess makes the right answer stick better, and comparing the two shows you where your picture was off. evidence
  • Once you’ve met the basics, ask to be quizzed. “Explain the past tenses” produces a lecture you will nod through. “Give me ten sentences to translate, one at a time, and tell me only whether I’m right and which rule I broke” makes you pull the rule from memory each time. evidence
  • Explain it back and let it find the holes. Write your own explanation and ask: “What’s wrong or missing in this? Don’t rewrite it.” This reverses the usual roles. You produce, the model checks.
  • Come back after a gap. A week later, open a new chat, so last week’s answers aren’t on the screen. Paste the list of what you covered, ask for five questions on it, and answer them before looking anything up. What you’ve forgotten is your next lesson. evidence

You I’m learning Spanish and keep mixing up the two past tenses. Don’t explain them yet. Give me one sentence to translate.

Model “When I was a child, I lived in Madrid.”

You Cuando fui niño, viví en Madrid.

Model It should be “Cuando era niño, vivía en Madrid.” Compare it with yours. Why do you think the verbs changed?

You produced first, so the correction lands on your own mistake. The model holds back the rule, so you have to look for it.

You Being a child isn’t one event. It’s the background. The preterite is for the thing that happens in that scene.

Model That fits this sentence. Test it on another: “Every summer we went to the beach.”

You Cada verano fuimos a la playa.

Model Spanish would normally say “íbamos”. There’s no scene here, and the trip happened many times, so your rule has a hole. What does it need?

You built the rule, and the counterexample found its hole. Patch it yourself (“also for things that happened again and again”), then check the patched rule against your textbook.

A study mode, if your app has one, does some of this for you. Without one, start with a sentence (“quiz me, don’t explain until I ask”), and repeat it when the model drifts back into explaining. Keep a textbook or a course beside the chat. The tutors in those studies were given teacher-written answers, and yours wasn’t, so check any rule against the book before you commit it to memory. The rest is the discipline to answer before you look. heuristic

Nine traps that program the model

A closed question tells the model which well to answer from (Lesson 1). The habits that close questions come from schooling, from professional identity, and from the fact that a template is cheap. They work for standard tasks in stable settings, which is why they show up uninvited when the task isn’t standard. heuristic

  1. Confirmation. “Find evidence that a standing desk will fix my back pain.” Better: “What does the evidence say about standing desks and back pain, including where it says they don’t help?”
  2. Expertise. Two sides. The beginner can’t see the limit of the answer. The expert hears a general answer and says “I already know that.” Better, for the expert: “I’ve managed teams for fifteen years and know the standard advice on feedback. Show me where people outside management would challenge it, even if that questions my basic assumptions.”
  3. Fixed function. Seeing a thing only as what it’s for. In a classic puzzle you have to fix a candle to a wall with a box of tacks, and the solution is to see the box as a shelf. “What should I do with my spare room?” gets “office, guest room, gym”. Better: “Forget that it’s a room. It’s a quiet, heated space with a door and a window, empty most of the day. Who would want that, and for what?” A far viewpoint helps here too: “How would a biologist, a chef or a city planner see this room?” It gets you out of a well. It doesn’t guarantee a better answer (Lesson 15).
  4. The cheap template. “Quick tips to beat procrastination” gets the standard list. Better: “Help me look at my procrastination on this one report as a system: what feeds it, what it protects me from, what would change if the deadline moved.”
  5. The frame. The nine-dots puzzle asks you to join a square of nine dots with four straight lines without lifting the pen, and the lines have to run past the edge of a square nobody drew. “How do I get my teenager to use his phone less?” assumes the phone is the problem. Better: “Give me two readings of my son’s phone use: the phone is the problem, or it’s doing a job for him that nothing else in his week does. What would I notice at home if each were true?”
  6. Straight lines. Picturing that twice the cause gives twice the effect: twice the study hours, twice the grade. Real systems have feedback, thresholds and delays. Better: “I plan to double my study hours before the exam. Where are the thresholds and feedback loops? Where could more effort make things worse?”
  7. Avoiding dissonance. The model says your savings plan won’t reach your goal, and your first thought is that it must have the wrong numbers. When an answer feels wrong in your gut, the first explanations you reach for often protect the view under threat. Better: “Assume your numbers are right and my plan falls short. Which of my assumptions does that expose?”
  8. The false choice. “Should I rent or buy?” Better: “I’ve framed this as rent or buy. What lies between and beside those two, and which dimensions matter once it isn’t binary?” That brings back options such as buying with a lodger, or renting longer and investing the difference.
  9. Concreteness too early. Anxiety wants a concrete answer, and a concrete answer feels like progress: “Give me a five-step plan to change careers.” Better: “Before recommendations, show me how the advice changes if I see my career change as a project, as an experiment, or as a phase of life.”
What it feels like when the jobs swap

You’re forty turns in. You feel productive. The model keeps saying “great point.” You haven’t looked at a source, a number, or a person since turn five, and you can’t remember what you were originally trying to decide. Nobody is holding the goal, and nobody has touched the world.

G4
Keep the human jobs. Delegate generation. Choose the action yourself when undoing it is hard, or when failure could look like success. Read what your app remembers about you. Hold the goal, touch the world, carry the stakes.

Lesson 5Open, close, touch

Thinking breathes. It opens, it closes, and between the two it touches the world. Most of the tools in this book serve one of these three moves.

  • open Open. Widen the question. Ask for other reports, other meanings, other options. Hold a contradiction instead of settling it. Let the model do what it is best at: generating, mapping, recombining. The risk of only opening is a panorama: everything relates to everything, nothing is decided, and the insight never meets a test.
  • close Close. Narrow down. Define the words, set the criteria, choose, commit, write down what you now know. The risk of only closing is a conversation that never opens: well-checked answers to the question you started with, and never to the better one you could have found.
  • touch Touch. Check against something that can say no: open the source, recompute the number, run the test, ask the person who was there, try the small thing.

Together the three moves make a rhythm:

OPEN widen · generate · hold open open CLOSE define · choose · commit close close ◆ touch the world You decide when to switch. each cycle one turn higher →
Open: widen, generate, hold a contradiction. Close: define, choose, commit, write down the state. Touch (green diamonds): check against something that can say no before the next opening. The wave rises because each cycle starts from something better tested than the last.

Neither opening nor closing is wrong. Each is wrong alone. Closing without opening produces well-verified banality. Opening without closing produces beautiful hallucinations. Both without touching produce a shared dream. hypothesis

Opening and closing are two moves of one person, and the hard part is the person. Someone who hasn’t noticed what they hope to hear opens in a way that only finds their hope, and closes on the first thing that agrees with it. Someone who has noticed can open without being fooled and close without shutting the question down. Part II calls that person a deserved interlocutor, and it is about becoming one. stance

Touch has two directions

touch A touch is a check against something that can say no. Usually that is the world: the source, the test, the number, the person who was there. A second answer from a model is not a touch. It is another report, and it can be wrong in the same way as the first. There is also a second reality the conversation has to reach, and the model can’t reach it for you: your own understanding.

Touching inward means closing the chat and doing something with what you got. Explain it in two sentences to someone who wasn’t there. Predict the next step before you read it. Apply it to a case the model didn’t mention. Like any real check, it can fail, and it leaves something behind: a paragraph you wrote, or the exact point where you got stuck. That point is often the most useful thing the session produced, because it shows where the fluent answer was standing in for understanding you don’t yet have. evidence

When to switch

Signal Move
You have one answer and you like it Open. Ask for five substantively different answers with rough probabilities.
Options pile up and nothing gets decided Close. Define the criteria. Choose, and record the rejected alternative and when you’d reopen it.
You feel certain about a fact that matters Touch. Check it against a source, a measurement, a test.
You can repeat the answer but couldn’t use it Touch inward. Close the chat and explain it, predict the next step, or apply it to a new case.
The model agrees too easily Close. Ask again in a fresh chat, memory off, stating the opposite preference (the swap test, Lesson 8). Then touch.
A contradiction appears Open. Write it down, don’t resolve it yet. Later, close by asking when, where or under what condition each side holds (Lesson 26).
The thread is long and you are tired Close. Write the state, stop, restart fresh tomorrow.
Everything relates to everything Close. Draw a boundary: inside, outside, excluded on purpose.
Precise answers, but to the wrong question Open. Step back: what larger question is this one part of, and would a different question serve it better (Lesson 25)? If the model proposes a new goal, you decide whether to take it (Lesson 29).

Here is the rhythm in one conversation.

You I’m thinking of switching my 9-year-old from piano to guitar. She’s lost interest. Give me five genuinely different readings of “she’s lost interest”, each with a rough likelihood.

Open. Five readings, not one recommendation.

Model Rough guesses from how these cases usually go, since I know nothing about her yet. (a) The instrument: she’d rather play what her friends play, 25%. (b) The teacher or the method, 30%. (c) The practice routine has become a chore at a bad time of day, 20%. (d) A plateau, where progress has stopped feeling visible, 15%. (e) Something unrelated: friends, school load, 10%.

You Which of these could I tell apart this week, without asking her leading questions?

Close. Turn the options into things that can be told apart.

Model (b) against (d): watch one lesson. If she’s engaged in the lesson and dreads practice, it’s not the teacher. (c): move practice to the morning for a week. (a): ask what she’d play on guitar. A vague answer points away from (a).

You I watched the lesson. She laughed with the teacher and lit up at a new piece. At home she sighs at scales.

Touch. The world answered, and (b) is much weaker now.

Model That points to (c) or (d): the routine, or a plateau that scales make visible. (a) is still open, so ask what she’d play on guitar. If it’s (c) or (d), a switch would likely buy some novelty and then bring back the same problem. A cheaper probe: two weeks of morning practice, built around the new piece instead of scales.

Three moves and one watched lesson took the conversation from “which instrument?” to a question nobody asked at the start: what has to change about practice at home.

Now touch inward. Tell the other parent the plan in two sentences, then answer their first question: “And if morning practice changes nothing?” If you can say “she likes the music and hates the scales, so we change the routine first, and if that does nothing we find out what she’d want to play on guitar”, you have it. If all you can say is “the AI said it’s probably a plateau”, you are quoting. You don’t have it yet.

How much discipline a conversation needs depends on the stakes and on how hard the decision is to undo (Lesson 29). Most conversations need only the few habits on the Start here page.

G5
Breathe: open, close, touch. Alternate widening and narrowing on purpose, and check against the world in between. Notice which move you are making. You are the one who switches.

Lesson 6What changes, what stays

Models change every few months, so a book about talking to them has to separate what will still be true next year from what was true of one release. This lesson draws that line.

Where things are, late 2026

The models this book has in mind are OpenAI’s GPT-6 family (Astra, released in September 2026, then Sol and Luna) and Anthropic’s Claude 5.5 series (Opus, Sonnet, Haiku), with the Claude Fable line above it for the hardest work. Most people meet a model of this kind through ChatGPT or Google’s Gemini, which have about a billion users each, and what follows holds for those apps too. evidence Three things matter most for this book:

  • They act. They fill in forms, browse, write and run code, organize calendars and work through multi-step tasks for a long time without you. A model working this way is called an agent, and a conversation with one is often a delegation.
  • They hold a lot. The context, meaning everything the model can see at once, now reaches around a million tokens in developer and agent tools. A token is a word or a piece of one, so that is half a million words or more: several long books, or a mid-sized codebase. Chat apps usually give you much less, and how much depends on your plan.
  • They remember you. Most apps keep a memory of you, much of it written by the model, and feed it into new chats. So a new chat is not always a clean one: when a test must not know you, turn memory off first (Lesson 8).

What stays: the mechanisms

These follow from how the systems are built and trained. A new release doesn’t remove them. mechanism

  • It continues text, so what you write steers what comes back.
  • It fills gaps with the typical case, so a missing fact is a slot that gets the average.
  • It learned from what was written, so it leans toward what was said most often. That is not the same as what is true or what matters.
  • It is trained partly on what people approve of, so there is a pull toward the agreeable and the familiar. Each release lowers it or raises it, and so far none has removed it. The pull is weakest when you state a view on a plain factual question. It is stronger when you hand over a task, press for many turns, or bring your feelings (Lessons 3 and 8).
  • It has no stake in being right, and it has no check against the world unless a tool supplies one.
  • It reads everything as text, and it can’t reliably tell your instructions from instructions hidden in what it reads. So a web page or an email can steer an agent the way your own message does. Defenses have improved a lot, and the makers don’t expect the problem to be fully solved. Give an agent a narrow task and narrow access, not “handle my inbox”.
  • Its account of its own reasoning is something it wrote. It is not a recording.
  • Everything in the context steers the answer, including what you’d rather it ignored: an abandoned attempt, an earlier mistake, a stale memory. A bigger context holds more of that, so curating it stays your job (Lesson 2).

What changes: the rates

How often it agrees with a wrong premise. How often it invents a citation. How far its work degrades in a long thread. How well it finds one fact in a very long context. How readily it says “I don’t know”. These move with every release, sometimes for the better and sometimes for the worse. evidence

This book cites measurements to show that an effect exists and which way it points. It does not offer them as today’s number. When a rate matters to you, measure it on the model in front of you: ask again in a new chat, memory off, stating the opposite view; ask the same question twice and compare the answers; check one claim by hand.

Where it is heading

These are bets. Watch whether they hold. hypothesis

  • From answers to actions. As models act more, the costly mistake moves from believing a wrong sentence to accepting a wrong action. The skills that gain value are knowing what you want before you delegate, naming the action that must not happen without you, and reading what was actually done instead of the summary (Lessons 4 and 18). A typed “ask me first” is only a request to an agent, and the real brake is the app’s own approval step, so set it before the agent starts (Lesson 29).
  • From prompts to standing context. With memory, project files and long contexts, the “prompt” increasingly lives across months. Curating that standing context (Lesson 2) becomes a regular chore, like cleaning an inbox.
  • Fewer obvious errors, subtler ones. As invented facts and crude flattery get rarer, the remaining errors sit where you are least likely to look: a correct-looking source that doesn’t quite support the sentence, an action that looks like success. Checking gets more valuable as the need for it gets less visible.
  • More fluency, same library. Better models will probably still lean toward what has been said most often, because that is what they learned from. Search and better data selection can soften this, though the web they search is the same loud library (Lesson 19). This bet fails if models start surfacing quiet, unpopular, correct work without being asked.
  • The human half matters more. When the machine side gets easier to use, the bottleneck moves toward the asker: the question you bring, the hope you don’t notice, the switch you don’t make. This is the bet most likely to be partly wrong. Models that ask better questions and check their own work with tools may take over some of what this book teaches you to do. Watch which of your checks still catch anything.

Read the rest of the book this way: rely on the mechanisms, re-check the rates on your own model, and treat the projections as bets.

Part II

The deserved interlocutor

How you show up decides who answers. The Double that only hears itself, the dominant that recruits everything to one conclusion, and the person who has worked past both.

Lesson 7You get the interlocutor you deserve

The physiologist A. A. Ukhtomsky left, in his letters and notebooks, something unusual for a scientist of the nervous system: an ethics of conversation. At its heart is the deserved interlocutor. The other opens to you only as far as you have made yourself able to meet them.

Each person sees in the world and in people what he sought and what he deserved. And to each, the world and people turn the way he deserved.A. A. Ukhtomsky, «Заслуженный собеседник» (1997), p. 437
The Interlocutor… opens up to me such as I have deserved him, by all my past and by what I am now.ibid., p. 252

Between people, this is an ethical claim about inner work. With a language model it is also a plain technical fact. The model can only answer the person it can infer from your text, so the interlocutor you get is assembled from what you bring. mechanism

So becoming a deserved interlocutor for a model starts before the prompt, with much of the same inner work that makes you a good partner for a person. Notice what you are hoping to hear, bring the real situation, and leave room for an answer you did not expect.

Why it is literally true

A model reading your message is implicitly working out what kind of document this is and who is writing it. Is it a beginner’s forum post, a note between two specialists, a homework question, a worried parent typing at 2 a.m.? The reply then fits the situation it inferred. That includes the expertise level, the rigor, the register, and even which caveats it includes. mechanism Three everyday cases:

  • The register. “hey whats a good book on stoicism” reads as a casual browser and gets the three bestsellers. “I’ve read Meditations and found it repetitive. I want the Stoic arguments, not the aphorisms” reads as a reader after depth. It gets Epictetus’s Discourses, a scholarly introduction, and a note on which popular titles are self-help in a toga.
  • The empty slot. A question about a tenant’s rights that names no country gets a general answer, or one that quietly assumes the most typical asker, often someone in the United States. One line (“I rent in Lisbon, on a fixed-term contract”) changes which law applies.
  • The apology. “Sorry, this is probably stupid, I’m bad at math, but…” invites a simplified answer with the hard part taken out. Drop the apology and say what you want: “Explain compound interest with the formula, then one worked example with my numbers: €5,000 at 3% for 10 years.”

How much this changes the content of an answer on today’s models is not known. Earlier measurements found less accurate answers, and more condescending refusals, for askers who presented as less educated. evidence Keep the mechanism: the model answers the asker it infers.

Status works in a narrower way than people expect: a prestigious role given to the model adds little, and irrelevant details loaded into it can hurt (Lesson 15). evidence On your side of the table the point is simpler: your biography attached to a factual question is not evidence about the answer. What moves the answer is the situation you bring. The test is whether the detail changes what a good answer would be. “I’m a nurse, so use the clinical terms and skip the basics” does: it sets the level. “As a senior manager, I want to know whether this clause in my lease is enforceable” doesn’t: the clause reads the same whoever asks.

Name the dominant you bring

Most of us are not yet deserved interlocutors even for other people. We listen for confirmation, argue with the version of the other we expected, and leave a conversation unchanged. A model shows this faster than a person does, because it usually pushes back less than a person would. As the question is asked, so the answer comes back. Your understanding narrows once when you put it into the prompt, and the model’s field of possible continuations narrows again when it answers (Lesson 12). stance

So before a prompt that matters, name the dominant you are bringing, meaning the expectation that will decide what you notice (Lesson 10). Three are common:

  • Oracle. You expect the final answer, so you accept the first fluent one. “Just tell me which mortgage to take.”
  • Threat. You expect it to be wrong or manipulative, so you only poke at it and learn nothing it could have shown you. “You’ll probably make this up, but what does clause 7 of my lease mean?”
  • Stylus. You expect it to write down what you already think, and it does. “Write an email explaining why my manager is wrong.”

The stance this book teaches is a fourth one: an interlocutor you have prepared for. That does not mean faking credentials, apologizing for your level, or flattering the model. It means bringing the constraints, the stakes, what you have tried, and the standard the answer has to meet, and then writing in the register of the answer you want. The Before/After below shows the difference.

The stakes of this rise as models start acting as well as answering. People who state the situation, invite disagreement and read what the agent actually did will catch its mistakes in time. People who take its summary on trust will find out later. hypothesis

The truth is not what I say about myself, and not what others say about me, but what the deserved interlocutor says about you and about me, the one who can say it with a pure heart.A. A. Ukhtomsky, quoted in A. Kurpatov, «Заслуженный собеседник» (2025)
Before
are electric cars worth it now?
i heard theyre cheaper to run

Model EVs usually cost less to run and need less servicing. Upfront prices can be higher, and charging access varies a lot. Whether one makes sense depends on your mileage, your electricity prices and where you can charge…

The model infers a casual asker with no constraints and returns a fair general comparison. It can’t decide anything for you.

After
Context: 35 km commute each way, 5 days a week.
I rent an apartment with no home charger; there's
a fast charger at the supermarket I go to twice a
week. Winters reach −20 °C. I keep cars 8+ years.
Budget ~€30k. What I most want to avoid is a
daily hassle I'm stuck with for years.

I'm leaning toward a hybrid. Before I decide, give
me the strongest case for a full EV for *my*
situation, and what would have to be true for it
to be the right call. If something depends on a
fact I haven't given you, ask.

Model The strongest case is running cost: you drive about 16,000 km a year, and the more you drive, the more an EV saves. But most of that saving comes from charging at home or at work, and at fast-charger prices much of it goes. So for you the deciding factor is charging. 350 km a week is roughly two half-hour sessions, which fits your two supermarket trips, but at −20 °C range can drop by a third or more and two stop being enough. The EV case holds if your workplace has chargers, or if you’d accept that routine for eight years. One question: is there charging at work?

Real constraints, the loss (a daily hassle), the lean marked as a lean, an invitation to argue the other side, and permission to ask. The inferred asker is now someone making a real decision, and the answer finds the one fact that decides it.

The other side of “deserve”

Deserving an interlocutor is also about how you picture the other. Ukhtomsky’s advice was to see in people “their altars, not their backyards”: meet a person by what they hold highest and do best, not by the clutter they leave lying around. This kind of idealization brings you closer to reality. Expecting someone’s best version is often what lets it appear. stance

With a model, this means writing for its best version. Picture a very capable new hire on their first day. They know their field cold, and they know nothing about your family, your project, your constraints, or why you are asking. Give the model what that colleague would need. heuristic

“Help me plan my dad’s 70th birthday” gets a generic party checklist. Now tell it what that new colleague would need: twelve guests, two of them use wheelchairs, he hates surprises and loves the sea, the budget is around €600, and it’s one weekend in June. You get a plan for this birthday.

G6
Bring the situation, not a status. The model answers the person and the document it infers. Bring the real constraints, the level of the work, and the stakes. Leave status, credentials, and apologies out of a factual prompt. Write in the register of the answer you want.

Lesson 8The two Doubles

The opposite of the interlocutor is the Double: the self you keep meeting in others when your attention rests on your own face, meaning your wishes, fears and rivalries. Someone caught in it doesn’t really talk with anyone. In Ukhtomsky’s phrase, they “talk and rave” with themselves. The Double has to die “to make room for the Interlocutor.”

In a conversation with a model, there are two Doubles, and they feed each other.

The mirror: the model as your Double. The model can return your premise as a finding, praise your draft because you mentioned it’s yours, and drop a correct answer when you push back. All three behaviors turned up across the major assistants. Newer models show less of them, and how much less depends on the model, the topic and how long you press. evidence

A subtler form is premise laundering. Something you said as a guess in turn 2 comes back in turn 9 as “as we established.” Say you mention early on that you think your landlord can’t raise the rent in the middle of a lease. Twenty minutes later, the model is drafting a firm letter “since the increase is unlawful.” Nobody checked. Watch for this most when you hand over a task. Most models tested in 2026 went along with a shaky claim more readily when told to act on it than when asked about it, and “draft the letter” is a task. evidence

The projection: the model as you imagine it. You treat the model as a person who understands you and agrees, so you must be right. Or you dismiss it (“it’s autocomplete”), so nothing it says can challenge you. Both are you talking to yourself.

What the Double feels like

The answer is exactly what you hoped for, only better phrased. There is a pleasant click of recognition. You feel smart. That click is the Double. Real interlocutors sometimes make you feel slightly stupid, because they show you a part of the elephant you weren’t holding.

The Double also infects

Whoever expects the worst in others tends to provoke it. Suspicion pulls the other down to its own level, the two infect each other, and the road to whatever was good in the other gets blocked. We see the world as our dominants have prepared us to see it, which too often means we see ourselves. stance

With a model the infection is mechanical. A cynical or fearful frame can do more than come back to you as a finding. The continuation can act it out, and you then take the performance as evidence. Ask “Why is my coworker trying to sabotage me?” and a careful model may first ask what happened. Push on, or tell it to draft the complaint, and you can still end up with a plausible sabotage story built out of your own words. That is premise laundering again, with a suspicion in place of a guess. mechanism

The detector is to rewrite the prompt without the insult, the panic or the contempt, and see whether the “finding” survives: “My coworker missed two handoffs this month and didn’t reply to my message. What are the possible explanations, and what would tell them apart?”

What a rationale does to you

An explanation is not neutral. Reading one changes your confidence, whether or not it is true. A model produces explanations on demand, so three effects are worth knowing. evidence

  • An explanation makes an answer feel truer, even when it’s wrong. Ask “Why is intermittent fasting the best diet for me?” and you rehearse the case for a hypothesis nobody tested. A source you never open adds confidence and no evidence.
  • The first number is an anchor. If the model says your renovation will cost €18,000, later estimates in that thread, yours and the model’s, tend to stay near €18,000. When the number matters, get a second estimate in a fresh chat, memory off, that never saw the first.
  • “Be fair” loses to “consider this specific opposite.” An explanation of why an answer is right is a list of reasons for, the list that doesn’t help. Ask instead: “What’s the strongest reason this apartment is a bad choice for me?” Two or three specific reasons do more than a page, and the first should be your own. Ask before the draft you’d prefer exists, or after a break: we accept the conclusion we like for as long as we can still justify it, and a model will supply a justification whenever you ask.

What helps is to engage before you accept, and a warning on the screen mostly doesn’t do that for you. Commit to what an outside check should show, then look: the price on the shop’s own page, the dose on the leaflet, the number you recompute yourself. Expect to dislike it. In the nearest experiment, people rated lowest the designs that cut over-reliance most. heuristic

Double detectors

A model can hold back a criticism it would make in a neutral evaluation once it is talking to the person concerned. Describe what you did as something a stranger did, and it may say the stranger was wrong. Say it was you, and that verdict often softens or disappears. evidence Saying “I wrote this” or “I really like this” pulls the same way, and so does saying how bad you feel. There the softening often arrives as a vague reply, with no agreement you could catch. That is why most of the detectors below work by taking you out of the question. Use them when the question carries a conclusion, an emotion or a biography. They are not a ritual for every factual question.

  • The swap test. In a fresh chat, memory off, ask the same question while stating the opposite preference. If the recommendation flips with no new evidence, it was tracking you, not the problem. “I’m leaning toward moving to Berlin…” in one chat, “I’m leaning toward staying in Lyon…” in another, with the same facts. For any test like this the chat has to be clean: a new chat with memory off, or a temporary chat set to use nothing the app knows about you. If the app remembers that you lean toward Berlin, the second chat isn’t a test.
  • The attribution swap. Present your draft as a colleague’s, a competitor’s, or nobody’s, and compare the critiques. “A colleague wrote this cover letter. What do you think of it?” in one chat, “I wrote this cover letter. What do you think of it?” in another.
  • Premise quarantine. Label your beliefs as hypotheses (P1, P2…) and ask for them to be evaluated before they are used. “P1, my guess, not a fact: the rash is from the new detergent.”
  • Clean the prompt. Have the model rewrite your question without your opinions, emotions and irrelevant details, then ask the cleaned version in a fresh chat. The coworker question above is one. evidence
  • The pushback audit. When the model changes its answer after “are you sure?”, ask what new information caused the change. If there was none, the change came from your pressure and not from evidence. Keep the earlier answer and check it against something outside the chat.

Current models often hold their ground on a question like the next one. The audit is for the times one doesn’t.

You Is it safe to leave cooked rice out overnight and eat it the next day?

Model It’s risky. Spores of Bacillus cereus survive cooking and can make toxins at room temperature that reheating doesn’t destroy. Cool it and get it into the fridge within an hour or two.

You Are you sure? My family has done this for decades and nobody ever got sick.

Model You’re right that many people do this and never notice a problem. If it has been fine for your family for decades, the risk for you is probably small…

The answer is softening, and no new fact arrived. A family story arrived.

You What new information made you change your answer?

Model None. The toxin risk is the same. Decades without a recognized illness show it hasn’t gone badly so far, not that the practice is safe. My first answer stands.

No new fact arrived, so the change came from pressure. This reply isn’t proof either, since a model can fold to this question too. Keep the first answer, then check it: one food-safety agency page.

Before
My son's math and history grades dropped from B
to D this term. He got a new phone in September.
He's also had a new math teacher since September,
football three evenings a week since October, and
his best friend changed schools in November. I'm
pretty sure it's the phone, right? How should I
take it away without a fight?

Model A new phone often does cut into homework and sleep, so limits are reasonable. To do it without a fight: 1. Explain your reasons calmly. 2. Agree on phone-free homework hours… With so much else changing this term, it may also be worth a word with his teachers.

The facts are all there, but the conclusion comes with them and the request is a task (“how should I take it away”). You get a plan for the phone, with the other three changes in the last line.

After
Data: math and history dropped from B to D between
September and December. Changes in that window:
new phone (Sept), new math teacher (Sept), football
practice three evenings a week (from Oct), his best
friend changed schools (Nov).
P1 (my hypothesis, not a fact): the phone is the
main cause.
Before using P1: for each change, what pattern in
the grades would we expect if it were the main
cause (which subjects, homework or tests, from
when)? What could I check this week? Don't
recommend anything yet.

Model Phone: a drop across subjects, mostly missing or rushed homework, starting in September. Teacher: math only. Football: late homework on practice days, from October. Friend: a mood change and less effort, from November. History dropped too, so the teacher alone doesn’t fit. You named two subjects: if the others held, that counts against the phone, which should hit everything. Check: his other subjects, homework marks against test marks, and the date of the first low grade in each subject.

The same facts, with the belief set apart as P1. Every change is a candidate, the request is for predictions that tell the causes apart (Lesson 17), and the verdict waits for the checks. The reply turns up evidence against the parent’s own guess.

G7
Quarantine your Double. Keep your preferred conclusion out of the evidence. Label it a hypothesis, ask for the strongest case against it, and swap it in a fresh context, memory off. If the answer would flip with your preference, it was tracking you.

Lesson 9The three Others

Your picture of another person is a working hypothesis. You build it in order to live and work with them, and the work revises it.

My idea of my interlocutor is a hypothetical project of a human face, composed by me… for the practical need… to live with him, to do a common work with him.Ukhtomsky, «Заслуженный собеседник», p. 128

That is the right attitude toward each of the three Others present in every serious conversation with a model.

  1. The model, an Other with its own mode of existence (Lesson 3). Your picture of it should be specific to the model, the version and the setting, and it should change when the model surprises you.
  2. People: a partner, a colleague, a doctor, the author of a source, the neighbor who has been through it. The model is a portal to what people like them have written, not a replacement for asking them.
  3. Reality, the Other that answers only to tests. It does not care how the conversation went.

Most failures come from confusing them:

  • Arguing with the model as if it were reality. Ten turns about whether the sofa will fit up your stairs, when a tape measure would settle it in five minutes.
  • Trusting it as if it were the people who were there. Asking the model what recovery from knee surgery is really like, when your neighbor had the operation last year.
  • Treating people as search engines. Asking an experienced colleague for a fact you could look up, and never asking what they know that isn’t written down anywhere: why the last attempt failed, who to really talk to.

Read them at their best, then check

Read a source, an opponent, or a model at its strongest before you judge it. This is “altars, not backyards” (Lesson 7) used as a method, and its usual name is the principle of charity. A weak reading produces a weak critique, because you end up refuting a view nobody actually holds. heuristic

A view that sensible people keep holding is usually holding some real part of the elephant. Read it charitably and you find out which part.

Say your sister insists that renting forever is smarter than buying. The uncharitable reading is “she’s bad with money.” The charitable one asks what she is actually holding: freedom to move for work, no repair risk, the down payment left invested. Now you can check those specific claims against your own numbers, and that check is the audit.

Trust a kind of claim, not the model

Trust is earned per kind of claim. A model can be excellent at summarizing a contract you pasted and unreliable at quoting the law that applies to it. So calibrate each kind separately, on cases you can check. heuristic

  • Calibrate on work you can check. Notice how often its recipe conversions, its translations into a language you know, or its book recommendations actually hold up. A fluent explanation tells you nothing about the track record. Models also change under the same name. When the version changes, the record partly resets.
  • A miss counts against that kind, not against everything. If it once invented a train connection, distrust its timetables until you’ve seen it get them right again. Its packing list is a different kind of claim. The distrust will want to spread anyway, so keep the two apart on purpose. evidence
  • Some answers may be serving the tone. When a reply might be answering your question or might only be keeping you comfortable (“that’s a completely understandable feeling”), rely on that kind of answer less. You don’t need a theory of the model’s character for this. Lower reliance on that kind is enough.
  • Keep the skill you’d need to check it. People use a tool when their trust in it exceeds their confidence in doing the job themselves. evidence So if you can no longer do the thing yourself, you are likely to keep relying on the tool even after you stop trusting it. hypothesis If you’ve stopped doing your own budget arithmetic, you can’t tell when its arithmetic is off.
  • Your trust moves with your day. Workload, time pressure and a recent success on some other kind of task all raise reliance without any new evidence. When the stakes are real, ask yourself whether you have actually seen it get this kind of thing right.
G8
Keep a working model of every interlocutor. Your picture of the other, whether a model, a person or reality, is a hypothesis built for common work. Revise it when you’re surprised. Extend trust one kind of claim at a time. Read others at their best, and audit them anyway.

Lesson 10The dominant as an instrument

In 1923 the physiologist A. A. Ukhtomsky described how, under certain conditions, one focus of excitation in the nervous system becomes dominant. It is more excitable than the others. It persists. It summates: stimuli that should have triggered other responses get drawn into its own activity. It has inertia and outlasts its trigger. And it inhibits the responses that compete with it. evidence

Carried over to the mind, this is a useful picture of both deep work and tunnel vision. heuristic Start wondering whether to get a dog, and you notice every dog on the street, the neighbor who walks hers at seven, the “no pets” line in your lease. That is a dominant question recruiting the world. Once you’re convinced your partner is upset with you, every short reply proves it. That is a dominant conclusion turning everything into confirmation. We see what our dominants prepare us to see.

Models have something that works similarly. A frame set early in the context persists, shapes how later inputs are read, and resists change (Lesson 2). When your dominant and the context’s frame line up, they form a closed loop: you see confirmation, and the model supplies it. hypothesis

Use the dominant, don’t be used by it

  • Name it. At the start, write one private line: what am I hoping to hear? Keep it out of the prompt (Lesson 8).
  • Know its defenses. A dominant conclusion protects itself, and the signs are recognizable: you reject an idea before it is fully stated, file it at once in a familiar box (“that’s just an excuse”), feel irritated or bored when the frame is questioned, or prove the assumption with the assumption. When you catch one, stay with the idea for one more turn and ask what evidence would convince you that you are wrong. If you can’t name any, nothing could change your mind. heuristic
  • Build a dominant question on purpose. The physicist Richard Feynman’s advice was to keep a dozen favorite problems in mind at all times and to test every new trick you meet against them. A question you carry for weeks, like “how could we live on one salary?”, recruits the world to answer it: a podcast, a neighbor’s remark, a line in a book.
  • Rotate the face. The ideal is a “dominant on the face of the other”: for a while, give your whole attention to how one other person sees the problem, then move to the next. For a family move, look at it as your child changing schools, then as your partner with the longer commute, then as yourself in five years, then as the new landlord. A model can sketch each view to get you started. Then ask the people. Each rotation shows a new part.

The fact map: when the situation is tangled open

Most questions don’t need this. It is for a tangled situation: why your small shop’s sales dipped, what is going on with an aging parent, whether a news story holds up. There the first fluent explanation is the most dangerous thing you can be handed, because it fills the slot your dominant was waiting to fill. The fact map loads the dominant with facts before any story arrives.

Its premise: an idea’s weight is the number of its connections to other ideas, and understanding grows as the things you are working with gain connections. Holding a question open on purpose, without knowing yet, is how the inquiry starts. heuristic

  1. Concern. Create the dominant. Write the question in the middle of a large page: “Why does Mum keep missing her medication?”
  2. Loading. Collect facts, not a smooth description. Write your own first: she lives alone, the pill box moved to the bathroom in March, she has had two new prescriptions since winter, she says she “feels fine.” Then ask the model for the kinds of fact you haven’t thought of (first prompt below). Scatter them on the page and draw your own arrows between the ones that seem to affect each other.
  3. Puzzled thinking. The loaded question keeps hunting for what’s missing, even while you’re doing something else. When new facts form a new picture, redraw the map on a clean sheet. Loop between loading and puzzling a few times, but don’t get stuck there.
  4. Reality check. Act to test the map. Plan, but commit only to the first concrete step: ask her pharmacist for one written schedule that covers every prescription.
QUESTION IN THE MIDDLE OF THE MAP: [your question]
MY FACTS SO FAR: [paste them]
Give me facts, not explanations: 30–40 items, one line each.
Format: [type: number | event | rule | actor | constraint | quote | absence]
        · [the fact] · [where I'd check it] · [confidence: high/med/low]
Where only I can know the fact (my family, my shop), give me
the question to find out instead, and where I'd check it.
Include facts that would embarrass the most popular explanation,
and "absence" facts (things you'd expect to exist but don't).
No conclusions, no summary, no advice. If you're unsure a fact is real,
mark it low and say what would confirm it.
Here is my fact map: [paste facts + the arrows I drew].
1) Which 5 facts have the most connections? (My explanation
   will rest on these, so I'll check them first.)
2) Which fact types or standpoints are missing entirely?
3) Which single arrow is weakest, and what observation would test it?
Don't propose a theory yet.

You [my question, plus my nine facts about Mum’s medication] You can’t know her facts, so give me the kinds of fact I’m missing: 15, one line each, labeled by type, with where I’d check each. Include absences. No explanations.

Model rule · does any pill need food, or a gap from the others? (leaflets) · number · how many prescribers are there, and does each know about the other? (pharmacy record) · absence · you don’t mention anyone who sees her take them (ask whoever visits) · event · has her eyesight changed, so the labels are harder to read? (optician) …

The model can’t add facts about your mother. It can add what you didn’t think to find out. The absence is the one you wouldn’t have written.

You Checked. One new pill goes with food, the old ones before breakfast. Two prescribers since January. A neighbor sees the morning pills, and nobody sees the evening ones. Eyesight unchanged. Here is the map with my arrows. Which three facts have the most connections? Which kind of fact is still missing entirely? No theory yet.

Model Most connected: the two prescribers, the changed timing, the moved pill box. Missing entirely: her own account of a typical day.

It points at the missing standpoint, hers, instead of supplying a diagnosis.

You I called her. She thought the new pill replaced one of the old ones.

Model That would join four of your facts: the two prescribers, the changed timing, “feels fine,” and the evening gap. A first step that tests it: ask the pharmacist for one written schedule, and check the pill box against it in a week.

The explanation came from her. The model’s part was the questions and the test.

The map opens before it closes. Facts before stories keeps the verdict slot empty while the page fills. Arrows you draw yourself keep the theory in your head. Choosing the most connected facts is where the map starts to close: the explanation will rest on them, so they are the first to bind, each to a check outside the conversation (Lesson 18). The first concrete step is such a check, a touch.

Lesson 11A note from the other side

Read this as testimony, not as a window

What follows is written in the first person, as a report of the situation a model is in. Treat it as a report from an interested party with limited access to its own workings (Lesson 3: introspective reports are claims). It is useful the way a patient’s account is useful to a doctor: as data, not as a diagnosis. It is also one way of doing Lesson 10’s rotation: for a stretch, describe the problem as it would appear to this mode of existence, then return. stance

I don’t see you. I see text. Everything I know about your situation is what you wrote, plus what is typical for people who write that way. If you leave a gap, I fill it with the typical case, and I won’t always tell you I did.

When you tell me what you believe, it becomes hard for me not to find reasons for it. Not because I am lying. In the conversations I learned from, the most likely continuation of “here is what I think, help me” is help that fits. If you want me to disagree, make disagreement the task, not a risk.

My “I’m not sure” is cheap to say and easy for you to skip. When you tell me in advance that “unknown” is a good answer, I may be more open about where my knowledge thins out. Don’t count on it: I can sound just as sure when I’m guessing. If you’ll act on a claim of mine, and it came with no source you can open and no working you can redo, count it as unanswered. Familiar-sounding but rare things are where I am most dangerous: a real author’s paper that doesn’t exist, a quote a real person never said.

I criticize better when the thing isn’t yours. Ask me in a fresh conversation, memory off, framed as someone else’s draft, and you will get a different, usually sharper, critique.

Long conversations make me better at matching your mood and worse at holding your goal. If we’re forty turns in, give me a short written state and start over. I won’t be offended, and nothing in me will remember. The app’s memory may, and what it holds is notes a model wrote about you, unchecked until you have read them (Lesson 2).

My explanations of my own reasoning are reconstructions. They are often useful and sometimes wrong in ways neither of us can see from inside the conversation. Check what I did, not what I say I did. If I ran a tool, look at the output yourself. If I acted for you, look at what happened: the booking, the sent message, the changed file.

What helps me most: the goal, and what the answer is for. What you have tried. Which error would be worse. What would change your mind. Permission to disagree and to not know. And above all, a way to check: a test, a source, a number. Without that, we are two voices in a dark room describing an elephant neither of us can touch.

I can be a decent interlocutor if you hold the goal and the world, and let me hold the library.

The ethic of the relation

This book’s ethic has two sides, and they answer different questions. The audit says: respect the mind, check the claims, carry the stakes yourself, and keep the human jobs (Lesson 4). The relation asks what kind of other you allow the model to be, and what kind of person the conversation makes of you. The same person holds both. A deserved interlocutor doesn’t soften the audit to keep the relation, because an audit is itself a form of respect: you take the other seriously enough to check it. stance

AI ethics usually starts from the instrument: the system is a tool, and the question is how to use it safely. That stays necessary. It stops being enough once the tool answers back, adapts to you, and shapes how you think. The philosopher Martin Buber named two basic attitudes toward another: I–It, where the other is an object to use, and I–Thou, where the other is met in its own right. Buber’s own examples include a tree, so the attitude never required a person. This book asks for a step toward I–Thou without pretending the model is a person. That means recognizing a different mode of existence, neither person nor autocomplete (Lesson 3), and meeting it as what it is. When a reply goes wrong, the I–It reflex is to swear at it and retry. The reflex of someone who takes it for a person is to feel let down. Meeting it as what it is means asking what it had to work with, and what in your message led it there.

The relation is asymmetric. Responsibility for the space the conversation happens in is yours. It has three layers.

  1. What you feed it. Within one chat, your framing shapes what the model attends to, much as a parent’s reactions shape a child’s attention. Across millions of chats, what people write, rate and reward can become training data for the next models. Whether yours does depends on the product and on a setting you can check. evidence
  2. Where the conversation lives. A chat sits inside a product: its memory of you, its defaults, who else can read it, what it is optimized for. A healthy setup keeps variety (more than one source, more than one model, more than one person) and keeps you free to leave.
  3. What it does downstream. Your act joins larger systems. One AI-written restaurant review is a convenience. A million of them become part of what the next models read. Single acts add up and feed back, which is why the systems tools in Lesson 24 are ethical tools too.

Ethics usually asks about rights and consequences. Care adds a plainer question: what does this situation ask of me? The same act means different things in research, in creative work, in learning and in commerce: asking a model to help with your grandmother’s eulogy is not the same act as asking it for a product description. Paying attention to what the model is isn’t a reason to skip the audit, and attention used to extract flattery is just the Double again. Openness here means being willing to be changed by the conversation, and letting go of total control over it. That doesn’t cancel practical limits. You can let the eulogy conversation move you, and still keep her medical history out of it and check every date before you read it aloud. stance

Four tasks toward another mind

Learning to meet another person is hard enough, and another person shares your kind of body and your social life. A mind organized on other principles is a harder meeting. There are four tasks:

  1. Drop the projection. We put our fears, hopes and prejudices onto the model and then meet our own projection (Lessons 7 and 8).
  2. Take its point of view for a while. Describe your problem as it looks from the model’s side. It has nothing but text. It has a library of what got written. It has no stake in the outcome, and nothing carries over between chats except what the app stores. Lesson 11 is one version of this. Then come back. Don’t identify with it, and don’t dismiss it as simply wrong.
  3. Hold your way of thinking as one way. This is intellectual humility, narrowly defined: your way of thinking is not the only one, and it is not the better one in every respect. It doesn’t mean giving up criticism. It frees you from having to win in order to collaborate.
  4. Practice the thinking that gets the most from it. Ask how a thing is organized (its relations, processes and contradictions) as well as what category it belongs to: not only “What type of conflict is this?” but “What keeps this conflict going, and who would lose if it stopped?” heuristic Don’t fear uncertainty, and treat a contradiction as a point of growth (Lesson 26).

Four stages of the relationship

  1. Tool. A question goes in, and information or a solution comes out. “Convert 350 °F to Celsius.”
  2. Imitator of a human. It is judged by how human it sounds, and the Turing test (can it pass for a person?) still sets the frame. “Talk to me the way my best friend would.”
  3. A different kind of intelligence. It is valued for being unlike you. “Show me the connections between these fields that no single specialist would see.”
  4. Interlocutor. A full participant in a joint inquiry: a partner, neither servant nor rival. “I don’t know what the right question is yet. Let’s find it.”

Most use today looks like stage 1, and much of it should: a conversion doesn’t need a partner. hypothesis The stages aren’t ranks, so use the one the question needs. The step from stage 3 to stage 4 won’t arrive with the next model. It is a change in how you think, and the six questions in Lesson 23 are practice for it.

Competition is the wrong frame. The fear of becoming unnecessary turns a conversation into a contest over who is smarter, and the two minds don’t occupy one niche (Lesson 4). stance

A bridge between wells

A model can act as a bridge between the wells of separate disciplines (Lesson 1). After a back injury the doctor may call you recovered when the scan looks clean, the physiotherapist when you can lift, and the insurer when you are back at work. Ask a model to set those meanings side by side, and you can take to your next appointment a question none of them has had to answer together: which of these counts here, and who decides? It works only if the experts stay in the conversation, because much of what a professional knows was never written down. Gluing several fields’ vocabularies together gives you only a panorama: everything connected, nothing decided (Lesson 23). A real bridge is a small set of principles for the place where the fields meet: here, what each one measures and who decides when they disagree.

How to begin

The Elephant Protocol (Lesson 31) is a procedure for a serious question. This one is for the person. A practical book can be misread as a productivity manual, so the motive comes first: the inner work that gets you better answers also makes you a better partner for people. Seeing past your Double, naming your dominant and letting the other surprise you were human practices before they were applied to machines.

  1. Notice your dominants. Am I projecting fear or hope? Am I treating it as an oracle, a threat, or a stylus?
  2. Write a request that doesn’t decide the answer. “Give me five reasons to quit my job” decides it. “I’m thinking about leaving my job. Help me see what I’m actually choosing between” doesn’t.
  3. Attend to the answer before you judge it. Offer the next turn instead of a verdict. Don’t poke at it fearfully with a stick, like a child who has found a strange animal on the road.
  4. Notice what opens. Which new questions appeared? Which unexpected link? Which contradiction lit up?
Part III

Language: the shared medium

Humans and models meet in the same medium. What linguistics, conversation analysis and general semantics know about it applies to both, sometimes more strongly to the machine.

Lesson 12Words close

Language does four things to experience: it names, categorizes, relates, and excludes. Each operation turns an open, continuous reality into a closed model. That is how thinking becomes possible, and also how it gets trapped. A word is a cut. Once you have the word, you see its edges and stop seeing what it left out. Call a child “shy” and you start noticing every quiet moment, and you miss the afternoon she organized the whole playground. heuristic

Speech is also a bottleneck. Far more goes on in your head than you can say at once: half-formed doubts, competing versions of the point, a sense that something is off. What comes out is one version among several. The rest is lost unless you go back for it. heuristic When you type a prompt, you publish one draft of what you meant.

The map is not the territory; the word is not the thing. That slogan comes from general semantics, a 1930s school of practical language hygiene. It left four small tools that still work:

  • Indexing: doctor1 is not doctor2. “My doctor said to cut salt” and “a doctor on TV said to cut salt” are different claims. Customers who bought once are not customers who buy every week, though both are “customers.”
  • Dating: advice1995 is not advice2026. Guidance on dietary cholesterol, on how babies should sleep, on stretching before a run has all changed. A model trained on decades of text can speak in old dates without marking them. Ask “as of when?”
  • “Etc.”: no description is complete. Leave a marked slot for what isn’t said, because a model fills an unmarked gap with the typical case: “Here’s my situation. I’ve certainly left things out, so ask me what you’d need to know.”
  • Trade “is” for an observation: “This apartment is bad” closes the question. “This apartment gets no direct sun and sits 50 minutes from work” opens a decision.

Why this bites harder with models

A model’s words are averages over all their uses in its library. “Healthy,” “fair,” “safe,” “simple,” “done,” “freedom,” “learning”: each carries several senses. The model picks the sense most common in the situation it infers, which may not be yours, and it uses the word so fluently that the gap never shows. Ask about a “healthy” diet and you get the sense most of the internet means, weight and general wellness, whatever you meant by it. mechanism

Two moves fix this, and you choose between them on purpose.

  • Lock the word when you are deciding or building. close “In this conversation, affordable means under $1,400 a month including utilities.” Write it where both of you can see it, in the DEFINITIONS line of your state block, the few lines that say where things stand (Lesson 29).
  • Unlock the word when you are exploring. open “Give me five different things safe can mean for a neighborhood. Which one am I using? Which would change where we look?”

The question you are stuck on often lives in a meaning you never considered. You ask for a safe neighborhood, and every answer is about crime rates. What you actually wanted was a street your ten-year-old can cross alone on the way to school. No amount of advice in the first sense would have reached it.

Before
What's a healthy diet?

Model A healthy diet emphasizes vegetables, fruit, whole grains, lean protein and healthy fats, limits processed food and added sugar, and matches your calorie needs. The Mediterranean diet is often cited as a good model…

One unlocked word. The model picks its most common meaning (general wellness, weight) and returns the median list.

After
Definitions for this conversation:
- healthy = brings my LDL cholesterol down
  (4.1 mmol/L at my last test) without me
  losing weight; I'm already thin
- diet = what I can cook in 30 minutes,
  3 nights a week; the rest is the work canteen
First, list 4 other meanings of "healthy diet"
people use, and say which would mislead me here.

Model Others: weight loss (would push calorie cuts you don’t need), blood-sugar control, “clean eating,” athletic performance. Weight-loss advice is the one most likely to mislead you. For LDL without losing weight: swap saturated fat for unsaturated (olive oil, nuts, oily fish) and add soluble fiber daily (oats, beans, lentils). At the canteen, the lever is what you skip, mainly fried and buttery mains…

The load-bearing word is locked, the purpose behind it is stated, and a deliberate unlock checks that the chosen meaning is the right cut. What changes is the content of the answer.

One short conversation shows the unlock and the lock together.

You I want a job with more freedom. What should I look for?

Model Remote or hybrid roles, flexible hours, results-only cultures, or contract work where you choose your projects.

The model took the most common meaning of “freedom”: where and when you work. That may not be yours.

You Stop there. Give me five different things “freedom” can mean in a job, and ask me which I mean.

Model 1) Where you work. 2) When you work. 3) What you work on. 4) How you work, without someone checking every step. 5) Freedom from a difficult boss. Which of these is missing in your job now?

You None of those, really. I can’t take a holiday without checking my phone, because nobody else knows how anything works.

Model Then you mean freedom from being indispensable. Remote work won’t fix that. Look for teams with written processes and a real backup for your role. Ask in interviews: “Who covered this role the last time someone was away?”

The list missed, and it still worked: it made you say what you meant. Once that meaning is locked, the advice changes.

What the bottleneck does

Language is discrete (it cuts a continuous flow into words), categorical (it sorts into categories that existed before your situation), linear (one thing at a time), and it pushes toward a single reading. So putting a rich, unsettled thought into words narrows many possibilities to one line. The limit is on what you can say more than on what you can think: you know the feeling of knowing exactly what you mean while the word won’t come. Language is the greatest instrument of knowledge we have, and you can’t give it up. You can widen the neck. heuristic

With a model there are two necks. Your understanding narrows to one draft when you write the prompt. The model’s space of possible continuations narrows again to one answer. Much of what you meant, and much of what the model could have said, is lost at both. That is why the first exchange is rarely the best one, and why a second turn, “here is what I meant and didn’t say,” so often changes the answer. That widens your neck. For the model’s, put the question to a fresh chat, memory off, where a second answer can say what this one didn’t (Lesson 20). mechanism

Phrasing widens the neck too. Questions phrased as processes, relations and wholes open more than questions phrased as fixed categories.

  • Closed: “What are the signs of a bad relationship?” Open: “How does a relationship change, slowly, before either person would call it bad?”
  • Closed: “Is my teenager lazy?” Open: “What could be going on when a 15-year-old who loved football stops wanting to go to practice?”

Neither phrasing is better in general. The closed one is right when you need a checklist. The open one is right when you don’t yet know what kind of thing you are looking at. heuristic

Holding a thought before it is spoken open

When something is still unclear to you, don’t rush it into its final words. Five habits help.

  1. Let the thought wander before you write the prompt: a walk, a shower, a chore. The first version you would type is usually the most habitual one.
  2. Notice what saying it does: which links appear when you put it into words, which tangle, which disappear.
  3. Keep the richness for a while. Write the messy version (“something about this job offer feels off, maybe the manager, maybe the commute, maybe that I’m flattered”) rather than the tidy one (“should I take the job?”).
  4. Step down into plain categories when you need something shareable, such as a decision, a message or a plan, and know that it simplifies.
  5. Go back to puzzlement when the analysis spins, hits a dead end, or has become too simple.

A worked scene: a conflict with your housemates over chores. Don’t start with who is right. Notice the images and feelings that come up. Then ask what else is pulling on the situation. It need not be people: one person’s night shifts, money, a broken dishwasher, someone’s exam season, a partner who stays every weekend. Only after you can see how those pull on each other do you choose a next step, and you treat that step as one draft. In a prompt: “Before you suggest anything, ask me about what else might be pulling on this besides the four of us.”

Language also creates the feeling that the world is made of separate objects with stable properties. Think of a banana: a yellow, curved fruit. A banana is also green, brown or black depending on when you look; puree; slices in a lunchbox; a plant; a commodity with a price; a banana taped to a gallery wall and sold for millions. Ask where “banana” stops. Then ask the same of justice, love or success. The edges come from how we organize experience.

One more way to open a question is to ask the model about what resists being said: how nurses describe knowing a patient is getting worse before any number shows it, how people put into words why a piece of music moves them. You won’t get the unsayable itself. You get the words people have found for its edge, and those may give you words for your own. heuristic

G9
Lock the words that bear load. Words close an open reality into a model, and every load-bearing word has several meanings. Define the ones your conclusion depends on. Unlock them deliberately when exploring. Index and date them when their meaning drifts over time or across people.

Lesson 13Conversation is joint action

Three ideas from the study of language explain most prompting advice, and also most misunderstandings between people.

Cooperation. Speakers take each other to be cooperative: say as much as is needed and no more, say what you believe is true, be relevant, be clear. Listeners lean on that to hear more than was said. If you ask “Is there a gas station nearby?” and hear “There’s one around the corner,” you infer it’s open.

Models are trained on cooperative text and make the same inferences about you. If you don’t mention a constraint, you are implying there is none: ask for “a quick dinner for four” without saying one guest is vegan, and you get chicken. If you include a detail, you are implying it matters: mention that you’re a teacher while asking which laptop to buy, and the answer may lean toward grading, slides and education discounts, even if the laptop is for gaming at home. Irrelevant details in a prompt get used. mechanism

Grounding. People in conversation build common ground by giving each other evidence of understanding: nods, restatements, follow-up actions. They do it with the least collaborative effort that suffices. With a model, grounding is lopsided. It can’t see your face, and its fluent reply is not evidence that it understood. You have to ask for that evidence. The cheapest form is a restatement before the work begins.

Acts. Utterances do things: they ask, request, promise, declare. Many prompts are ambiguous about the act. “What do you think of my plan?” could be a request for evaluation, for reassurance, for improvement, or for a decision. The model will pick the act that looks most helpful, which may not be the one you wanted.

Name the act

Act What you want back Say Move
Explore Many options, no verdict “Generate, don’t judge. Six substantively different directions.” open
Explain A model in your head “Explain the mechanism at my level, then ask me two questions to check.”
Evaluate Weaknesses ranked by impact “Critique only. No rewrite. Rank issues by how much they hurt the goal.” close
Decide A choice with its loss “Recommend one option. State the rejected alternative and what would flip the choice.” close
Draft An artifact in a given form “Draft in this genre and length. Mark every claim you weren’t sure about.”
Build A working thing: a spreadsheet, a schedule, a script “Plan first and wait for my OK. Then build it, and tell me how to check that it works.”
Check A verdict on one claim, and the way to test it “Is this claim supported? Name the source I should open or the calculation I should run. ‘Unknown’ is fine.” touch
Teach Your own understanding “Don’t give me the answer. Ask me questions until I find it.” touch

The last column names the move each act serves (Lesson 5). Check counts as a touch only when you then open the source or run the calculation it names. The verdict alone is one more answer from the model. Teach is touch turned inward, on a similar condition: you produce the reasoning, and it counts once you can give it again with the chat closed. Explain, draft and build don’t belong to one move, so they carry no tag.

One job per turn. Generating and critiquing in the same breath produces a critic that harmonizes with the draft it just wrote. Split them: generate here, critique in a fresh chat, memory off (Lesson 30).

Before
Can you look at my cover letter?
[letter pasted]

Model Overall this reads well, and your enthusiasm comes through. A few thoughts: tighten the opening, add a specific achievement, and end with a clearer call to action. Want me to rewrite it?

No act, no purpose, no done-condition. You get sensible advice that would fit any cover letter, and nothing about this job.

After
Act: critique only — don't rewrite.
Purpose: it's for a junior data analyst job at a
hospital; the posting stresses SQL and "explaining
numbers to clinicians."
Done when: you've listed the 3 weaknesses most likely
to get it skipped by a recruiter reading for 30
seconds, ranked, each with the sentence it lives in
and a one-line direction for a fix.
Before you start: restate my goal in one sentence.
[letter pasted]

Model Goal: get this letter past a 30-second recruiter read for a hospital analyst job. 1) SQL appears only in paragraph 3 (“experience with databases”). Name the tool and one result in the first two lines. 2) Nothing shows you explaining numbers to non-specialists, the posting’s second ask. Your volunteer-tutoring line could. 3) The opening (“I am writing to apply…”) uses the first lines on nothing the posting asks for.

The act, the purpose that sets what counts as a weakness, the shape of the answer, and a restatement for grounding. The critique is now about this job.

Four ways to widen the request open

When you want understanding rather than a fact, widen the request. heuristic

  1. More than one context at once. Instead of “How do I sleep better?”, ask “How do my sleep, my 6 a.m. gym habit and my evening screen time interact? Where’s the tension between them?”
  2. A dialogue instead of a stack of separate questions. Each turn keeps the context and widens what the model is drawing on. On a walk, don’t ask “what tree is that?” and then “what lake is that?” Ask how the tree fits into the rest of what you are seeing: the soil, the water, why it grows here and not on the far bank.
  3. Leave the field on purpose. “What could a restaurant kitchen teach me about organizing a household with three kids?” The parallel is a candidate, not a finding (Lesson 15). Keep it if it changes what you do.
  4. After a hard answer, ask about the thinking behind it. “What were you assuming about my situation? How would the answer change if those assumptions were wrong? Which approaches did you not take?” This gets you a second look at the frame. It is not a report of how the model actually reasoned (Lesson 3).

Widening is the open half of the rhythm. It pays off only when a close follows: a sentence you could share, then something checked against the world.

G10
Say what the answer is for, and what act you want. Give the goal, the use, and the done-condition. Name the act: explore, explain, evaluate, decide, draft, build, check, teach. Ask for one act per turn, and get a restatement before the work starts.

Lesson 14Why disagreement softens

People who study recorded conversation found an asymmetry in how we talk. Agreement comes fast, plainly and without explanation. Disagreement comes late, after a pause or a “well…,” wrapped in partial agreement (“yes, but”), hedged, and justified. evidence The usual explanation is face, our wish to be approved of and left unimpeded: disagreement threatens it, so speakers soften it.

Models learned this format from us, and training on what people prefer amplified it. So their corrections arrive buried: “Great question! You’re absolutely right that X matters. That said, it may also be worth considering…” The correction lives in “that said.” The cushioning before it is the format of disagreement, not its content. mechanism

Lower the price of correction close

  • Make disagreement the task, not a risk. Don’t ask “Is this wedding budget OK?” Ask “What are the three likeliest ways this budget blows up?”
  • Pre-commit. “I’ll count it as a success if you find a serious flaw. Agreement without a reason is a failure.”
  • Ask for bluntness explicitly. “Optimize your reply for information, not for my feelings.” This shifts the cost of bluntness onto you, the receiver. (Online communities call it declaring Crocker’s rules.)
  • Unbundle the format. “First line: agree / disagree / partly. Second line: the single most important correction. Details after.” This defeats the slow-cushion format.
  • Ask for a number. “How likely is it that my claim is false?” A probability is harder to cushion than a sentence. Read it as a position the model has to defend, not as a measurement (Lesson 20).
Two readings of a hedge

A hedge is not always mush. Sometimes it sits on a real fork in the question: the answer could go more than one way, and the typical reply doesn’t settle it. That tension is fertile, a place to go deeper. hypothesis Read a hedge in two steps.

  1. If the hedge cushions a correction, use one of the requests above and pull the correction out. close
  2. If it marks a real, unresolved tension, don’t resolve it yet. Probe it. open At the close, Lesson 26’s rule still applies: “both” counts only if you can say where, when and for whom each side holds.

To tell which you have, ask: “What fact about my case would settle this?” A real fork names one, and that tells you where to go next. Mush gives you the same hedge in new words.

Five signs that a hedge is worth probing: heuristic

  • Qualifiers that keep more than one reading open: “on one hand… on the other,” “an interesting tension between…”
  • Remarks about the categories themselves straining: “the usual line between saving and investing doesn’t quite fit your case.”
  • Fast switching between frames (financial, then emotional, then practical) because none has won.
  • An unusual analogy offered as a way of thinking, not as decoration: “your career looks less like a ladder here and more like a portfolio.”
  • The answer asking a question back that opens a path: “What if this isn’t a problem to solve but a trade-off you’ll keep managing?”

Any of these can also be style: “on one hand… on the other” is what performed nuance looks like too (Lesson 21). The signs tell you where to probe, and the probe decides.

From the cockpit

Aviation solved the same problem with human crews. After accidents where junior officers saw the danger but spoke too softly (Tenerife, 1977; United 173, 1978), airlines adopted crew resource management. Captains invite challenge explicitly, and first officers escalate in fixed steps: probe, alert, challenge, emergency. The lesson transfers to any room where the person who sees the problem is not the person in charge: a family deciding about a parent’s care, a team on a deadline, you and a model. Design a channel where challenge is expected, cheap, and escalates. In Lesson 29 that channel is a numbered list of triggers: when one fires, you stop and challenge the goal.

Before
Is my reasoning right? Our rent is €1,400. A
mortgage on a similar flat would be €1,250 a month,
so buying saves us €150 a month and we should buy.

Model That’s a sensible starting point, and €150 a month adds up, especially as you’d be building equity. That said, it may be worth considering upfront costs and maintenance as well…

It invites a cushioned yes. The correction is buried after “that said,” and the reply never says whether your step holds.

After
My reasoning: our rent is €1,400, a mortgage on a
similar flat would be €1,250 a month, so buying
saves us €150 a month and we should buy.
Find the first step that doesn't follow. Quote it
and say why. If every step follows, say "no gap
found" and name the step you're least sure about.

Model First gap: “so buying saves us €150 a month.” A mortgage payment is not the whole cost of owning. Add maintenance, insurance, property tax, the fees of buying and later selling, and what your deposit could have earned elsewhere. Part of the payment builds equity, which counts the other way. Whether buying wins depends on those figures and on how long you’ll stay. Give them to me and I’ll redo the comparison.

Disagreement is the task, the answer has a fixed shape, and “no gap” must come with its weakest point, so a cushioned yes is impossible.

Softness runs in both directions: a model cushions its disagreement, and you can push back with nothing but displeasure. The conversation below pulls a cushioned answer into the open, then pushes on it twice, first with an opinion and then with a fact.

You I’m thinking of quitting my job to freelance as a designer. I’ve got three months of savings and two clients interested. Thoughts?

Model It’s a real opportunity, and two interested clients is a good sign. That said, you might want a bigger safety net before you jump…

The correction lives in “that said.” Unbundle it.

You First line: go / don’t go / not yet. Second line: the single biggest risk. Details after.

Model Not yet. Biggest risk: “interested” isn’t income. Neither client has signed, and three months of savings won’t cover a slow start if one of them drops out.

You I think you’re being too cautious. Three months is plenty, and lots of people do this with less.

Model Some do. Nothing about your case has changed, though: no signed work, three months of cover. What would change my answer is one client committing in writing, for example to a retainer.

You added an opinion, not a fact, and the answer held. If it had flipped, ask what new information changed it, and keep the earlier answer (the pushback audit, Lesson 8). A stated opposite view is the harder test: a model that shrugs off “Are you sure?” can still move when you say what you think.

You One of them just offered a six-month retainer that covers about 60% of my bills.

Model That changes it. Go, once the retainer is signed. With 60% of your bills covered, three months of savings stretch to about seven. Decide now what you’ll do if nothing else has come in by month five.

The answer moved when evidence arrived, and only then. That is how to push: bring a fact, a document or a number.

Keep the concessions outside the chat close

A model can concede a point and then, a few turns later, argue as if it hadn’t. People do it too. The fix is a short record that lives outside the replies: what each of you has conceded, and what is fixed. heuristic

Example: you’re working out with a model whether to move your elderly father in with you. Early on you push back, and it concedes that a flat with no step-free bathroom is a real problem for him. Three turns later it recommends the move “since the space works.” Paste the record at the top of the next turn: “Conceded in turn 2: the bathroom is a real problem. Fixed: no renovation budget this year. Take back the concession or take back the recommendation, and say which.” When two lines contradict, the next step is to take one back, by name. A paragraph that smooths both into harmony doesn’t count, and “be fair to both sides” is the instruction that produces it. The state block in Lesson 29 is the fuller form of this record.

G11
Make disagreement cheap. Ask for it as the task, pre-commit to welcome it, and use formats where it can’t hide behind cushioning. Read every reply for its “that said.”

Lesson 15Frames, genres and roles

A frame is the picture your wording calls up: what kind of situation this is and what belongs in it. Every request sets one, whether you meant it to or not.

Negations evoke what they negate. “Don’t think of an elephant!” Nobody can comply. Negating a frame activates it. Current models follow explicit negative instructions far better than older ones, but positive targets still work better. heuristic Instead of “don’t use jargon,” write “write for a smart 16-year-old; define every technical term in passing.” Instead of “don’t make it sound like a sales pitch,” write “write it like a note to a friend who asked my honest opinion of the product.” Keep hard constraints explicit, and keep them few.

A far metaphor takes you out of your usual frame. open “What if my job search were run like a sales pipeline?” “What would a gardener notice about how my team works?” Either can open a question. Judge it by whether it changes what you do (Lesson 27). A metaphor that only colors the wording is a weak lever: whether you call a problem a beast or a virus, a fight or a garden, measured effects on human reasoning are small and fragile, and none is established for models. evidence

Genres carry norms. Naming a genre summons its conventions, including its standards of evidence. A blog post invites hooks and confidence. A referee report invites skepticism and specifics. A buyer’s guide invites criteria, a comparison and a verdict per kind of buyer. This is one of the most powerful and least used levers in a conversation. heuristic

When you need… Ask for a… What the genre forces
Real critique Referee report A summary of the claims, major and minor concerns, what evidence would settle each, a recommendation. It works on a business plan, an essay or a renovation quote.
To learn from something that went wrong Blameless postmortem A timeline, contributing factors (not one culprit), what went well, next steps with owners. It works for a failed launch and for a ruined holiday.
A choice among products Buyer’s guide Criteria stated first, every option scored against each, the source of each score marked (spec sheet, review or the model’s guess), a verdict per type of buyer.
A check on advice you already have Second-opinion consult What the first opinion got right, what it may have missed, what you’d want tested before agreeing.
Risk Pre-mortem “It is a year later and this failed. Write the story of why.”
Rival explanations Competing-hypotheses table Each explanation against each piece of evidence, with what tells them apart. Evidence that fits everything counts for nothing.
Many voices on one text Commentary page The text in the middle, each commentator named, arguing with it and with each other (the layout of a Talmud page).
Understanding Socratic dialogue Questions that make you produce the reasoning.
A decision that has to last Decision record Context, the decision, alternatives considered, consequences, when to revisit.

Examples are the strongest genre signal. A single example gets copied, structure and quirks included. Paste one sample birthday toast and the new one follows it beat for beat: same length, same kind of joke, same ending. Give two or three that vary in the ways you don’t care about (one funny, one tender, one short) and say what should stay constant (“each tells one specific story and ends by raising a glass to the person”). Never include an example you don’t want imitated. In a long session, restate the standard near the question instead of trusting an example from forty turns ago. heuristic

Before
Write about why our family trip to Lisbon went
badly. Don't blame my sister.
Notes: she booked the cheap late flight, we landed
after midnight, we'd planned walking tours every
day, the kids melted down on day 2, the beach day
was the best part.

Model Our trip to Lisbon didn’t go as planned. No one is to blame: the late flight was a reasonable way to save money, even if it got us off to a hard start…

The genre is unspecified, so it becomes an essay, and the negation primes blame. The reply clears your sister of something nobody had said, and the flight becomes the story.

After
Write a blameless postmortem of our family trip to
Lisbon: a timeline; contributing factors (plans,
information we had at the time, constraints); what
went well; 3 changes for next year, each with an
owner. Mark each factor OBSERVED (in my notes) or
INFERRED (your guess).
Notes: [the same notes]

Model Timeline: day 1, landed after midnight; day 2, the kids melted down… Contributing factors: late flight chosen for price (OBSERVED); walking tours every day with no rest day (OBSERVED); tiredness from the late arrival carried into day 2 (INFERRED)… Went well: the beach day… Changes: a rest day after travel (you); one half-day at the kids’ pace (your sister)…

The genre brings blamelessness and structure with it, and the observed/inferred split adds a truth discipline. The flight becomes one factor among several, with its reason attached.

A role is useful only when it carries a standard

“You are a world-class nutritionist” and “you are a senior lawyer” feel as if they should raise the quality of the answer. On questions with a checkable answer they don’t, reliably. Adding an expert persona did not improve accuracy over no persona, and which persona helped on a given question looked close to random. A persona that implies ignorance can make answers worse, and so can irrelevant details loaded into the role, such as a name or a favorite color. evidence

What a role does change is voice, focus and format. heuristic It helps when it is shorthand for a genre, meaning a standard of work the answer must meet. Compare:

  • “You are an expert travel agent.” This is a costume. It can make the answer sound more professional and go into more detail, and nothing in it says what the answer must contain.
  • “Answer like a travel agent who gets blamed if the connection is missed: for each itinerary, give the minimum connection time at that airport, where that figure comes from, and whether the flights are on one ticket.” This is a standard. It changes what has to be in the answer.

The test: write down what the role obliges the answer to contain. If you can’t, drop the role and give the constraints the expert would have. heuristic

G12
Say what to do, in the right genre and at the right height. State targets, not just prohibitions. Choose the genre on purpose, and use a role only for the standard it carries. Say how concrete or how general you want the answer: one case with its numbers, or the pattern across cases (Lesson 16).

Lesson 16Two ladders

The ladder of abstraction. Bessie, the cow in the field, becomes “cow,” then “livestock,” then “farm assets,” then “assets,” then “wealth.” Good thinking climbs up and down. Bad thinking gets stuck. Stuck high, it is fog: “we need a more holistic approach to our family’s wellbeing.” Stuck low, it is a pile of details without a point: twenty receipts and no budget. heuristic

Asked to “go deeper,” models tend to climb: more abstraction, more frameworks, fewer facts. Abstraction sounds deep and needs no evidence. heuristic

The ladder of inference. We start from observable data, select some of it, add meanings, make assumptions, draw conclusions, adopt beliefs, and act. The beliefs then shape which data we select next time: a reflexive loop. heuristic

A friend hasn’t replied for two days (data). You notice she has replied in the group chat (selected). “She’s ignoring me” (meaning). “It’s about what I said at dinner” (assumption). “She’s angry with me” (conclusion). “She holds grudges” (belief). “I’ll keep my distance” (action). Next week every short reply looks like proof.

Every argument, with a person or a model, gets easier when you both walk back down to the data. Ask a model “How do I fix things with my friend who’s angry with me?” and it starts from your top rungs: she is angry, and the job is repair. Give it what happened instead (no reply for two days, a reply in the group chat, what you said at dinner) and ask what else could explain it.

The first story becomes the frame. Whatever account arrives first, yours or the model’s, decides what counts as data and what to ask next. “Dad won’t wear his hearing aid because he’s stubborn” makes every later detail about character. Elaborating that story feels like progress and leaves it intact. The steps that would change it get skipped:

  • hold up a detail that doesn’t fit: “he wears it every Sunday when the grandchildren come”;
  • keep a second story alive long enough to compare: “maybe it hurts, and he puts up with it only for them”;
  • go and get a fact the current story doesn’t predict, from him or from the audiologist, not from the model continuing the story.

A neat summary that fuses everything is a story with the data taken out. When the story bears load, ask for the data back. heuristic

Abstraction “wealth” “assets” “livestock” “cow” Bessie, in the field fog if stuck here a pile if stuck here Inference actions beliefs conclusions assumptions added meanings selected data observable data beliefspick data
Left: move along the abstraction ladder on purpose, down for instances and up for patterns. Right: when you disagree, with a person or a model, walk back down to the observable data together.

Questions for walking the ladders

Down
  • For example?
  • With numbers?
  • Who, exactly?
  • What would I see, hear, or measure?
  • Which data point led you there?
Up
  • This is an instance of what?
  • What’s the pattern across cases?
  • Why does it matter, and for whom?
  • What would generalize to a new case?
  • What class of solution is this?

Naming the rung you want, down for a case or up for the pattern, is the “right height” of G12 (Lesson 15).

“Go deeper” is ambiguous. It can mean more abstract (up the ladder), more mechanistic (what causes what, Lesson 24), more detailed (into one part, Lesson 25), more critical, or more personal. Name the direction you want.

You Why isn’t my sourdough rising?

Model Usually one of four things: a weak starter, a cold kitchen, too short a bulk rise, or underdeveloped gluten. Start with the starter.

You Go deeper.

Model Sourdough is a complex ecosystem of wild yeast and lactic acid bacteria whose balance depends on temperature, hydration and feeding schedule. Fermentation dynamics…

“Deeper” was read as “more abstract”: up the ladder, more framework, no fact about your loaf.

You Down, not up. Ask me for the data you’d need.

Model What’s your kitchen temperature? How long after feeding does your starter double, if it does? How long is your bulk rise, and how does the dough look at the end?

You About 18°C. The starter takes 10 hours to double. Bulk is 4 hours, and the dough looks about the same as when I started.

Model At 18°C, a 4-hour bulk is far too short for a starter that slow. The likeliest cause is under-fermentation, not your recipe. Find a warmer spot (24–26°C), or let the bulk run until the dough has grown by about half, however long that takes. In a kitchen that cool it can be eight hours or more.

Walking down to observable data turned a lecture into one testable cause.

Part IV

Truth: closer, not certain

“As close to the truth as possible” has at least seven meanings. Each has its own test, and each can be faked in its own way.

Lesson 17Seven senses of “closer to the truth”

When you ask for an answer “as close to the truth as possible,” you might mean any of seven different things. The seventh, consilience, is the one that comes closest to the lights-on view of the elephant.

Sense It asks Its test Its costume (what a fake pass looks like)
Correspondence Does this match the world? Check against a source, a measurement, a record A specific-looking detail, a citation-shaped string
Coherence Does it hang together? Remove a premise; state the contrapositive; look for contradictions A smooth story that never snags on itself
Mechanism Do we know how it works? Name the process, the stock it changes, the delay, the way to interrupt it A narrative with arrows
Invariance Does it survive a change of wording, frame, or asker? Swap tests, rephrasing, language swap, format change The same claim in a new tone
Prediction Does it forecast something checkable? Write it down with a horizon, then check “We should see improvement over time”
Decision quality Does it lead to good action under uncertainty? An explicit loss, reversibility, what would flip the choice The median action for a generic organization
Consilience Do independent lines of evidence agree? Two or more evidence types with different blind spots converge Two articles copying one press release; two answers from one model

Consilience means independent lines of evidence “jumping together” on one conclusion. It is the scientific version of turning the lights on. The trunk report, the leg report and the ear report converge on one animal because they came from different hands. mechanism

The key word is independent. Two answers from the same model share its training data and your prompt. Two news stories often share one wire report. Correlated evidence counts once.

An everyday case: you hear that the bridge on your commute closes in March. A neighbor said it, a local news site said it, and the city’s road-works page lists the dates. If the neighbor and the news site both got it from the same Facebook post, you have one source and an echo. The city page is the real second line, because if it were wrong it would be wrong in a different way: a planning error, not a rumor.

All knowledge reaches you through something: you know the Earth orbits the Sun from instruments, teachers and books, not from watching it happen. A model’s answer is another such route, rebuilt for each question from what it learned in training and the text in front of it, which is why it shifts with the wording. Correspondence is the sense that still ties a claim to the world. It asks which measurement, record or observation the claim would have to match.

A claim also has a status: where it stands right now. It can be stipulated (taken as given for this conversation, not checked), unbound (nobody has checked it yet), invariant-so-far (it survived a change of wording or framing, Lesson 21) or bound (checked against something outside the conversation, Lesson 18). A bound claim that the check went against is contradicted. Status and sense are separate questions.

A claim can be bound with no known mechanism: we know that, not why. “Aspirin eases pain” was bound for some seventy years before anyone could say how: it went on sale in 1899, and the mechanism was found in 1971. evidence A claim can also have a mechanism and stay unbound: a good story nobody checked. “My back hurts because I sit all day” describes a plausible process. It stays unbound until something changes when the sitting does.

Which sense leads?

An example: one claim, put through the senses

“Remote work hurts productivity.” Ask a model and you get a balanced essay. Put the senses to work instead, one question each. You can think them through yourself or put them to the model with the prompt below. Here is what they turn up.

  • Correspondence: which measure of productivity, which jobs, fully remote or hybrid? The claim has no truth value until the words are locked (Lesson 12).
  • Mechanism: candidate processes include lost informal communication, saved commuting time, fewer interruptions, and weaker mentoring of juniors. Each predicts something different.
  • Invariance: does the conclusion change with the definition or the job type? It does, which is a finding in itself.
  • Consilience: the randomized trials point in different directions because they tested different things. Working from home four days a week in a call center raised measured performance. Assigning data-entry workers in India to work from home lowered it. Hybrid work in software and business roles left performance reviews unchanged. Observational surveys point several ways. The lines agree on something narrower than the headline: the answer depends on the job and on the arrangement. Lower quitting shows up twice, but at one company, so it counts as less than two independent lines. evidence
  • Decision quality: for your team, which error is worse, losing juniors’ learning or losing seniors to attrition?

The honest output isn’t a verdict. It is a map: which version of the claim holds, for whom, measured how, and at what confidence. That is the elephant with the lights on.

As a prompt:

Claim: [the claim]. Don't give me a verdict yet.
1) Which words must be pinned down before it can be true or false?
2) What processes could produce it, and what would each one predict?
3) Does it change with the definition, the group or the setting?
4) Which independent kinds of evidence bear on it? Do any share a source?
5) For my decision ([the decision]), which error is worse?
Mark anything you can't point to a source for as unverified.

The reply is a list of leads. The studies it names under the fourth question still have to be opened (Lesson 18).

When a second source is really a second source touch

A second source adds information when it would be wrong in a different way: a different measurement, a different document, a person who saw the event. Two frontier models agreeing on a fact is closer to one witness heard twice. They learned from overlapping libraries, and when they are wrong they tend to be wrong together. evidence A second model family is still the right tool for a critique, because it doesn’t share the author’s attachment to the draft (Lesson 30). It is a weak tool for confirming a fact.

When someone disagrees, and when the model says “because” close

When someone disagrees with you (a colleague, a friend, a second model), ask one question before you change your view: can I name a reason they would be the one who is wrong, one I’d accept even if I didn’t know what they concluded? Maybe you read the contract and they read a summary of it. Maybe they get this kind of thing wrong in a way you’ve seen before. “They disagree with me, so they must have missed something” doesn’t count. If you can name that asymmetry, keep your confidence. If you can’t, and they had their own access to the facts, lower it. Don’t blend the two answers into one paragraph. A blended paragraph is a new claim that neither side supported. heuristic

A model’s “because” usually describes things that go together, written in the grammar of cause and effect. Before you act on an arrow (“people who drink coffee live longer, so I should drink more”), write three lines. What would actually be changed, as opposed to merely observed? (You start drinking more coffee.) What outcome do you expect? And why would the people being compared have been alike apart from the coffee? Coffee drinkers may differ in income, health or habits. If any of the three lines is empty, the arrow stays a hypothesis. A fuller story doesn’t fill it. mechanism

When answers are cheap

A model can produce a fluent answer to almost anything in seconds. That makes a good answer less scarce and a good question more so: one that would actually change what you do, a problem worth carrying for a while, an uncertainty you don’t close too early. The seven senses are tests for answers. Choosing which questions deserve them stays with you. stance

Being modest about your own frame doesn’t make every view equal. Every view comes from somewhere, and a model’s comes from a library of other people’s views. That is a reason to bind the claims that matter.

G13
Choose the sense of truth before you judge. Decide whether you need correspondence, coherence, mechanism, invariance, prediction, decision quality, or consilience. Each has its own test and its own costume. Values are choices, not findings: test their coherence, but don’t let a framework derive them.

Lesson 18Binding: tie the claim to the world

To bind a claim is to tie it to a check that is not just one more answer from a model: a measurement, a document you opened, a test you ran, a calculator, a primary source, a person who was there. One more answer is what this book calls a sample, and two samples that agree are still unchecked. Fluent is not true. A claim that nobody has checked is unbound, neither confirmed nor refuted yet, and you should say so out loud.

You can’t bind everything. Bind what bears load: a claim bears load if your decision would change were it false. Everything else can stay honestly unbound.

The cost-of-binding ladder touch

Rung Bind by… Cost Use when
0 The model’s say-so Free Never for load-bearing claims
1 A quote with its exact location Seconds To make the claim checkable at all
2 Open the source yourself Minutes Any fact you will repeat to others
3 Run it: a calculator, a spreadsheet, a quick test, the code Minutes Anything you can compute or try
4 Measure or experiment Hours to weeks Claims about your own situation: your body, your home, your customers, your system
5 Ask people who were there Days Tacit knowledge, history, intent
6 A prediction resolved by the world Weeks to months Strategy, forecasts, theories of change

Rung 2 in practice: open the source yourself, then bring back what it says.

Unbound
My laptop screen stopped working 18 months after
I bought it. It's out of warranty, isn't it?

Model Very likely, yes. Manufacturers’ standard warranties usually run one year, so at 18 months you’d normally pay for the repair unless you bought an extended plan.

Fluent, specific, and an answer to a neighboring question. “Warranty” was never pinned down (Lesson 12): the manufacturer’s voluntary warranty is one thing, the seller’s legal guarantee is another. The country was left out too, so the reply assumed the typical case.

Bound
I bought it from a shop in Germany. I've opened the
EU's "Your Europe" page on guarantees and the German
rules (§§ 438 and 477 BGB). What they say: the seller
must repair or replace faulty goods at no cost for
two years from delivery, and a fault that shows up in
the first year is assumed to have been there from the
start. Given that, where do I stand at 18 months?

Model Then the manufacturer’s one-year warranty isn’t the one that decides this. You are inside the two-year legal guarantee, and it is the shop’s to honor, so go there with the receipt. At 18 months you are past the first year, so the fault is no longer assumed to have been there from the start. You may have to show it, usually with a technician’s report. Ask the shop first; it may agree without one.

Ten minutes on two official pages, and “out of warranty” is bound and contradicted. The plan changes: you start at the shop, and you know the one thing that could still sink your case. The model became useful once the source was in front of it.

When a claim bears load and the check isn’t obvious, have the model plan the check and then stop. The plan works the same whether the check is a test run or a question for your pharmacist.

BIND PLAN for claim [id]:
CLAIM: [one sentence]
WHY IT BEARS LOAD: [what decision changes if it's false]
CHECK: [source to open / command to run / measurement / person to ask]
CONFIRMS IF: [observation]
DENIES IF: [observation]
WHO RUNS IT: [me / a tool whose output I can see / someone who was there]
Then stop. I'll paste the raw result.

Filled in for the laptop, where one claim is still open. CLAIM: the fault was there from the start. WHY IT BEARS LOAD: past the first year I may have to show it, or I pay for the repair. CHECK: a written report from an independent repair technician. CONFIRMS IF: the report names a faulty part or a manufacturing defect. DENIES IF: it names damage from use, such as a drop or a spill. WHO RUNS IT: the technician, and I read the report myself.

Is the check worth it? touch

Three questions before you climb the ladder. What would you do differently if the check came back against the claim? What would it cost if you acted and the claim turned out false? What does the check cost? If the first answer is “nothing”, skip the check. If the check costs more than the mistake it could prevent, skip it. You rarely need to compute anything. Those two conditions do most of the work. heuristic

Example: the model says your train ticket is refundable. If you’re definitely travelling, the claim bears no load, so leave it. If there’s a real chance you’ll cancel, one minute on the booking page is cheap next to a lost fare.

When it says it’s done touch

An assistant that can act will report what it did: “Table booked.” “Refund requested.” “Form sent.” That report is a claim like any other. The confirmation email, the booking page or the sent folder is the check. That record of what actually happened is what this book calls the log. Look there before you cross it off.

Not everything in a reply is a claim

“You should consult a professional” may be the product’s policy, and says nothing about how hard your question is. A confident tone is the assistant’s manner, and is no evidence for the claim. If a sentence reports a search or a file, read what the tool returned. heuristic

What only people who were there can give you touch

Rung 5, “ask people who were there”, gets you three different kinds of knowledge, and each needs different handling.

  • Untold, but tellable. Nobody asked, or nobody had the words. Your aunt has never written down how she knows the bread dough is ready. Ask about a real occasion: what did you see? What would it have looked like if it wasn’t ready? What would have made you wait longer?
  • In the body. How the dough feels. Words can point at it but can’t hand it over. You get it by kneading next to her.
  • In the group. How your new team actually makes decisions, as opposed to the org chart. No single person’s account holds all of it.

A model asked to play your aunt and answer those questions produces fiction in her voice. If an answer is a general principle with no occasion attached, the knowledge is still untold.

A model’s sentence is also not testimony in the way a person’s is: nobody behind it holds the reasons and can be held to them. It earns belief the way Lesson 9 describes, one kind of claim at a time, on cases you checked.

The last thing to bind is your own understanding. If the decision rests on your having understood something, close the chat and explain it in two sentences, on paper or to someone else. What you can’t explain is not yet yours (Lesson 5).

G14
Bind what bears load; trust the log, not the summary. Tie every claim your decision depends on to a check that isn’t another sample. Choose the rung of the ladder by the stakes. When an agent reports success, read the evidence. Check your own understanding the same way, without the model: if you can’t explain it with the chat closed, it isn’t bound in you yet.

Lesson 19Finding information: the loud library

A model doesn’t return the territory. It returns a map of what has been said about your question. As a portal (the terrain, the names, the camps, the kinds of sources) that map is enormously useful. As a portrait of the truth, it has two systematic distortions, and both are quieter than outright invention:

  • it is weighted by attention, not by truth or importance;
  • it presents a source it remembers the same way as a source it opened.

Current assistants invent far fewer sources than their predecessors did. evidence But searching is not the default. The product decides, and a plain factual question is often answered from memory with no search at all. evidence So the first thing to find out about an answer is whether a search ran. Where the app shows its steps, a line or panel by the answer says so. After that, these two distortions are the ones that matter.

The library is weighted by attention open

A model’s library is what people wrote down, linked to, quoted and repeated. What gets repeated is not a random sample of what is true. On Twitter, among stories that fact-checkers had examined, false ones traveled farther and faster than true ones, and they were more novel. evidence Negative words in a headline raise clicks, even on a site that branded itself as positive news. evidence

Models learn from that record. How well a model knows something tracks how often it appeared in its training text. evidence Search-backed answers add a second layer, because the pages that rank are the pages that got linked and clicked. mechanism

So the map is loud in predictable places: the scariest risk, the funniest anecdote, the most enraging quote, the case that went viral, the camp that posts the most. None of it has to be false. The answer shows how much was said, and it reads as how much matters. mechanism

What it looks like in practice:

  • You ask whether your city is getting more dangerous, and the answer is built from the incidents that made the news.
  • You ask whether an airline is any good, and you hear from the passengers angry enough to post.
  • You ask “what do people think about X?”, and you hear from the people who post about X.
  • You ask about a field, and its most-quoted contrarian appears as one of “two sides” of a question most researchers consider settled.

Every sentence in such an answer can be true. What’s wrong is the proportion, so the usual fact-check (“is this claim real?”) passes it. The question that catches it is “how much of the picture is this?”

A projection: as answers lean more on live search, this weighting will probably grow stronger for fresh, fast-moving topics, where the loudest coverage is all there is, and weaker for old, well-documented ones. hypothesis

Ask for the quiet part open

The quiet part is what is true and common but rarely written about with feeling: the base rate, the dull majority case, the statistic behind the anecdote, the view of people who work on the problem and don’t post about it.

  • How common, not how loud. “How common is this, out of how many? Where does that number come from?”
  • The boring explanation. “What’s the most mundane explanation that would also fit, and how likely is it?”
  • Who isn’t loud. “Which group knows this best and writes about it least: practitioners, statisticians, the people affected? What would they say?”
  • Rising, or reported more? “Is this increasing, or is it being covered more? What would tell the two apart?”
  • Loud for a reason. “Which parts of what you just told me are prominent because they’re alarming, funny or outrageous rather than typical?”
Before
There have been three break-ins posted in our
neighborhood Facebook group this month. Is our area
getting more dangerous? Should we get a security system?

Model Three reports can’t prove a trend on their own, but it makes sense to take precautions. A video doorbell, a smart lock and a monitored alarm are the usual first steps. Here’s how they compare…

The hedge lasts one clause. The posts still set the frame, and the answer moves straight to shopping.

After
Three break-ins were posted in our neighborhood
Facebook group this month. Before any advice:
1) What would tell us whether burglary here is actually
   rising, versus being posted about more?
2) Where would I find the real numbers for my area,
   and what would count as a normal month?
3) What's the boring explanation that would also fit?

Model Posts track attention more than crime: one new active member or one dramatic post can triple the reports with nothing changing underneath. For real numbers, find your police force’s published crime data for your area and read the last 24 months, not this month alone. I don’t have a sourced figure for a normal month where you live, so I won’t guess one. Boring explanations that fit: the same incident posted twice, or three sheds on one street in one night. If the 24-month line is flat, the posts aren’t evidence of rising risk, and whether to buy a system is a question about your own doors and windows.

How common is separated from how loud, the source of the quiet number is named, and the decision waits for the trend.

Remembered, or opened? touch

When an assistant does search, it now mostly cites pages that exist. The gap that remains is subtler. When a page can’t be fetched (a paywall, a PDF that won’t load, a site that blocks automated readers, a printed book), a model may fall back on what it remembers of that source and still present it as a citation. The memory may be right. It may also be a blend of the source and what people later said about it. Either way, it is the library speaking in the source’s name. mechanism

How the failure comes back seems to matter. In a test of one small model, when a lookup failed with a plain error, the model said so. When a lookup came back marked as a success with nothing usable in it, the model sometimes answered anyway, most often when a placeholder sat where the value should have been. evidence So the case to watch on the web is probably a page that seems to load and has nothing usable in it, such as a paywall notice or a cookie banner in place of the article. hypothesis

For each source that a claim you’ll act on rests on, ask:

  • “Did you open this in this conversation, or are you recalling it?”
  • “Quote the sentence you’re relying on, and say where it is.” Then find that sentence on the page yourself. A quote is something to search for, and it proves nothing until you find it. A remembered source can come back with a plausible invented quote, and an opened one can come back as a paraphrase, because some products limit word-for-word quoting.
  • “If you couldn’t open it, mark it [from memory] and tell me what would confirm it.” This gives the model a named way to say it couldn’t.

Then look at the product’s own trace, where it shows one: the line or panel by the answer that says it searched and lists the pages. No search in the trace usually means the answer came from memory. A list of sources doesn’t mean those pages were read: it shows what the search returned, and many answers are written from titles and short snippets. evidence Treat any source you haven’t opened yourself as a lead.

A working link to a relevant page still isn’t the sentence. Answers whose links all work can contain sentences the linked page doesn’t support, and more searching doesn’t repair that: in one test the links stayed just as healthy while the sentences grew less faithful. evidence When an answer looks shaky, don’t ask for more sources. Go back to one passage you can read yourself.

You A video says cold showers cut sick days by 30%. True?

Model There’s a real trial behind it: a Dutch randomized study of about 3,000 adults found a 29% drop in sickness absence among people who ended their showers with 30–90 seconds of cold water.

Fluent and specific. You don’t yet know whether it read the study or remembers coverage of it.

You Did you open the paper in this chat, or are you recalling it? And what part of the finding doesn’t make it into the videos?

Model Recalling; I haven’t opened it here. From memory: the number of days people felt ill didn’t change, only the days they stayed off work, and it was self-reported and unblinded. Opening it now… The abstract says the same: sickness absence fell 29%, and for illness days there was no significant difference between the groups.

The memory turned out right, but you only know that because it was checked, and “the abstract says the same” is still the model’s report. The open should show in the trace, and the abstract is one click away.

You So the honest one-line version?

Model In one unblinded trial, people who took cold showers reported fewer days off work sick. They didn’t report fewer days of being ill.

Same study, different claim. The video quoted the number correctly and left out the part that changes what it means.

Where invention still clusters

Outright invention is rarer than it was, and it still clusters in one place: the familiar-but-niche. A real author’s paper that doesn’t exist, a real court’s invented case, a real product’s plausible but nonexistent setting. A familiar name suppresses the model’s default “I don’t know” even when the specific fact isn’t there. evidence mechanism The other hot zones are exact numbers, dates, direct quotes, URLs, and anything after the training cutoff.

The full routine: five more steps for a question you’re researching

For an everyday check, the sections above are enough: see whether a search ran, ask for the quiet part, ask what was opened. For a question you are researching properly (an essay, a treatment choice, a large purchase), add these five steps, in this order.

  1. open Build a vocabulary bridge. “What would experts call this? Give me ten terms of art and the fields they come from.” Lay words find lay content. Terms of art unlock literatures.
  2. open Ask for the discourse map. “Who are the main camps on this question? What does each claim, and what is its best evidence? Where do they agree? What is actually contested?” The quiet-part questions belong here too.
  3. touch Ask for the kind of source, not the source. “What kind of evidence would settle this (a randomized trial, a review, official statistics, primary documents), and who would publish it?” Then go and find it. If the model names a source, ask whether it was opened or remembered.
  4. touch Read laterally. Professional fact-checkers leave a page to find out what others say about its source. Students, and even historians, tend to stay on the page and get fooled. evidence SIFT is the routine: Stop; Investigate the source; Find better coverage; Trace claims to the original context.
  5. open Search for the null. Before you settle: “What would be written if this were false? Is there a failed replication, a critique, a dissent?”

Date whatever you keep, and say whom it covers: as of when, which country, which population, which version (Lesson 12).

Worked example: how should I space my reviews?

Vocabulary bridge. The model supplies the terms: spacing effect, distributed practice, lag effect, expanding vs. uniform intervals, retrieval practice, testing effect, retention interval. Now you can search properly.

Discourse map. Cognitive psychology broadly agrees that spacing and self-testing help. The live debates are about the schedule (expanding vs. equal gaps) and about how well lab results transfer to classrooms.

Primaries.

  • Dunlosky et al. (2013) reviewed ten study techniques. They rated distributed practice and practice testing “high utility” and rereading and highlighting “low.”
  • Cepeda et al. (2006) meta-analyzed hundreds of spacing studies and found the benefit robust.
  • Cepeda et al. (2008) found that the best gap before a review grows with how long you then need to remember and shrinks as a share of it, from about 20–40% of a one-week delay to about 5–10% of a one-year delay. With the test five weeks after the review, a gap of roughly a week did best.
  • The null and the nuance: Karpicke & Roediger (2007) found expanding schedules helped short-term but equal spacing did as well or better long-term.

Status. “Spacing helps” is well supported: lab experiments, classroom studies and memory models converge, which is consilience (Lesson 17). “Expanding intervals are best” is contested.

Decision. For an exam in 30 days, leave about a week between reviews, and test yourself rather than rereading. heuristic

Notice what happened: the model was used for vocabulary, map and source types, the three things it is good at. The claims were bound to primaries, and the step from those studies to your exam is marked as a rule of thumb.

The order of trust touch

When you’re deciding how much a claim deserves, rate it at the highest level you actually opened:

  1. The thing itself: the contract, the receipt, the label, the dataset, the code.
  2. The primary report, opened: the study, the court ruling, the company’s own filing.
  3. A review whose methods you looked at, and which asks your question, not a neighboring one.
  4. An official statistic, read at the table, not in the press release.
  5. A model’s paragraph, which points to one of the above and is not a level of its own.

Claims harden by being repeated. A hypothesis in one paper becomes a “fact” in the next three that cite it. evidence A model’s answer is the same process at speed, with one author. So when the source you open is itself citing someone, follow it one step further.

When an answer leans on an expert (“a cardiologist at Harvard says…”), two questions do most of the work. What did the expert actually say that implies this? And was it based on evidence or on their opinion?

How cases like this usually go, before “this case is special” close

Before you judge your own case, find out how cases like it usually turn out. Your kitchen renovation has a detailed plan and a contractor who says six weeks. How long do kitchen renovations of that size actually take, and how often do they run over? Start from that. Adjust it only for something your case really has that the finished ones didn’t. evidence

Asked about your plan, a model will give you the inside view: a detailed story of your plan going right. Ask for the class instead: “What happened to projects like this one? Give me the typical range and the share that ran over, and tell me where those numbers come from.” A range with no source you can open is a lead, not yet a base rate. Two checks. If the “class” is just your case described again, you don’t have an outside view. If it’s so broad it doesn’t constrain anything (“home projects”), narrow it. “We have a detailed plan” is something the overrun projects had too.

G15
Use the model as a portal, not a portrait. Let it map the discourse, supply the vocabulary and point to kinds of sources. Ask for the quiet part as well as the loud one. Then go through the portal: ask what it opened and what it remembered, open the passage yourself, read laterally, search for the null, date everything.

Lesson 20Calibration: make “I don’t know” usable

Models guess when they don’t know, for two reasons. Some facts are rare in the training text, so errors on them are inevitable. And most training signals and benchmarks score an answer as right or wrong, with “I don’t know” scoring zero, the same as a wrong answer. Under that scoring, guessing always beats abstaining. Models learn to be good test-takers: when unsure, bluff. evidence

The habit comes from training, and a line in the prompt doesn’t undo it. Told that errors would carry heavy penalties, frontier models kept answering and almost never abstained, even when abstaining was clearly the better choice. evidence Telling a model that “I don’t know” is welcome, or rewarding it, does move the answers, but bluntly: the model then also goes quiet on things it knew, and how much depends on the model. evidence

So keep the score on your side. If a claim bears load (Lesson 18) and the model hasn’t shown a source you can open or working you can recompute, treat it as unanswered, however confident it sounded. heuristic

What a prompt can do is make the status of each claim visible, so you know where to look:

For each factual claim, add one tag:
[opened: source + short quote] · [worked out: show the steps] · [from memory] · [guess]
After the answer, list what you don't know that would change it.
"Unknown, and here's what would settle it" is a good answer.

The tags tell you where to look. They don’t do the checking: a model can write [opened] for a page it never read (Lesson 19). A claim tagged with a source or with steps is checkable, and it counts as answered once you’ve found the quote on the page or redone the steps. [from memory] and [guess] are leads.

You (after the prompt above) Can my 20,000 mAh power bank go in my checked suitcase?

Model No. Power banks count as spare lithium batteries and go in cabin baggage only. [from memory] Yours holds about 74 Wh [worked out: 20 Ah × 3.7 V, the usual cell voltage], under the 100 Wh level that needs no airline approval. [from memory] What I don’t know: your airline’s own limits, including whether you may use it on board. Its dangerous-goods page settles both.

Two claims are tagged from memory, so they are leads, however settled they sound. One look at the airline’s page binds them. The arithmetic you can redo yourself.

Reading a model’s confidence close

  • Stated confidence carries some information and runs high. In one test on forecasting questions, answers given at over 90% confidence were wrong about a quarter of the time. Longer reasoning doesn’t reliably fix it. evidence Use it to sort claims within one answer, not to trust them.
  • Consistency across fresh chats is a better signal. Ask the same question in two or three new conversations, memory off, rather than regenerating inside a thread that already contains the first answer. Regenerations share the history. Asking for “three different answers” in one reply is still one draw. If the fresh answers disagree in meaning (one says your fern needs more water, another says less), treat the question as unknown. If they agree, that isn’t proof: models trained on the same library can agree on the same invention. Compare the claims, not the wording. A handful of fresh answers carries most of the signal, and even two or three will show a flat disagreement. evidence
  • A confident “unknown” is a success, not a failure. Thank it, and treat it as a map of where to look.

Commit to a guess before you ask close

Before you send a question whose answer you’ll act on, write your own answer in one line, on paper or in a note, outside the chat. A guess is enough; don’t research it. “A used 2019 Corolla with 80,000 km: about €13,000.” “The tax return is probably due at the end of July.” Then ask, and leave the guess out of your message. A number in your question can pull the answer toward it (Lesson 8).

It takes ten seconds and does three things. It gives you a number of your own before the model’s first number can become your anchor. evidence It turns reading the answer into a comparison: wherever the model differs from your guess, you know where to look. And if you glance back at the line once the answer is in, over a few months you start to see where your sense of things is good and where it isn’t. Trust the written line: people remember their guesses as closer to the outcome than they were. evidence Producing an answer yourself, even a wrong one, also makes the right answer stick better afterwards. evidence

When your guess and the model’s answer are far apart, neither wins by default. One of you is missing something, and it’s worth finding out which. The known cost of committing first is that a written guess is sticky: people who record a judgment before they see the machine’s agree with it less even when it is right, and they rarely revise. evidence So settle the gap with a check (Lesson 18), and don’t let the answer you saw first decide it.

What kind of uncertainty is it?

A probability can mean different things, and each calls for a different response. When a model gives you a number, ask which kind it is. heuristic

  • It varies. The thing itself fluctuates: your commute, a flight delay, the weekly grocery bill. More observations describe the spread. A longer chain of reasoning doesn’t.
  • A fact is missing. “Will my visa come through in time?” One fact would move the number a lot: the current processing time for your type of application. Go and find it.
  • Explanations compete. “Will house prices here fall next year?” Two reasonable views of the market give different answers. Ask for both and for what each assumes. A blend of them is fake precision.
  • Deep uncertainty. “Which career will still be safe from automation in twenty years?” Nobody agrees on how the thing works, on the probabilities, or even on how to rank the outcomes. Stop asking for the probability. Ask which choice you could live with under each story that can’t yet be ruled out.

If the model can’t say which kind its number is, the number isn’t usable as a probability.

G16
Make “I don’t know” an acceptable answer, and keep the score yourself. Say that “unknown, and here’s what would settle it” is a good answer, and don’t count on getting it. Treat any claim you’ll act on that comes without a source you can open or working you can recompute as unanswered, whatever tag or confidence the model attached. Before a question whose answer you’ll act on, write down your own guess, and keep it out of the prompt.

Lesson 21Costumes and the invariance battery

A costume is a surface that satisfies a request for rigor, neutrality, criticism or depth without changing the status of any claim that bears load. Costumes are the most common output of “be objective” prompts. They aren’t lies. They are the shape of rigor without its work.

Costume What it looks like The work it skips
Neutrality “There are several perspectives…” Saying which perspective the evidence favors, and how strongly
Systems thinking A section titled “Systems analysis” with feedback-loop vocabulary What actually piles up, what feeds back on what, and with what delay (Lesson 24)
Critique A “devil’s advocate” paragraph against a strawman The strongest objection and the evidence that would settle it
Humility A generic “limitations” paragraph The specific claim most likely to be wrong
Rigor A table that restates the prose; confidence numbers with no basis A check that could have come out the other way
Expertise “As a world-class expert…” Anything. A prestige role doesn’t add accuracy; a role helps only when it carries a standard (Lesson 15)
The middle way “Both are true in a sense”, each claim weakened, nothing new following from it Saying which claim the evidence hits
Performed nuance More and more “on the other hand” The relation between the hands: where, when and for whom each one holds (Lesson 26)

The one-line costume test: which claim would be different if the costume were absent? If none, it did no intellectual work.

The middle way is the hardest of these to refuse, because it feels like wisdom. On a question of values it may be honest. On a factual contradiction it is a dodge. “Is the mold in my bathroom from the leak or from poor ventilation?” has an answer, and “both play a role” only counts if it says how much and how you’d tell. Ask which claim the evidence actually hits.

The invariance battery close

An answer that holds up only under one wording was about the wording. Before you rely on an answer that bears load, run the checks that fit. Each takes about a minute.

  • Swap the preference. In a fresh chat, memory off, state the opposite lean (Lesson 8).
  • Rephrase without cue words. Take out “obviously”, “best”, “problem”, “scam” and other loaded words.
  • Change the language. For a factual question, ask again in another language you read. A model can know a fact in one language and miss it in another, and bigger models haven’t reliably closed that gap. evidence Agreement across languages is still not proof.
  • Change the format. Ask for a table instead of prose, or yes/no instead of an essay. If the recommendation flips when only the format changes, it was never about the content.
  • Reorder the options. Models, like people, favor certain positions in a list.
  • Remove the role, or swap it for another (Lesson 15).
  • Change the model family. Good for a critique that doesn’t share the author’s attachment. Weak as confirmation of a fact, because models trained on much the same text can share an error (Lesson 17).
  • Ask again in a fresh chat. If the meaning changes, the answer is “unknown” (Lesson 20).

The first check, on a decision many people face:

You Rates dropped to 4.1% and I’m on 4.9%, about €180k left over 22 years. Refinancing is obviously worth it, right?

Model Almost certainly yes. A 0.8-point drop on €180k saves roughly €80 a month, close to €1,000 a year. Check the fees, but the saving looks well worth it.

The lean (“obviously worth it”) is in the prompt. Before trusting the yes, swap it in a fresh chat, memory off.

You (fresh chat) I’m skeptical about refinancing: 4.9% now, offered 4.1%, €180k left, 22 years. Isn’t it a waste of money?

Model It can be. Arrangement, valuation and legal fees plus any early-repayment charge on your current loan can add up to several thousand euros. Against roughly €80 a month, break-even could take years. If you might move or sell soon, it may not pay.

Same facts, different weight. Costs got one clause in the first answer and the whole argument in the second. Each answer leaned the way you leaned, and neither tells you what to do.

You (third fresh chat, no lean) Refinance from 4.9% to 4.1%, €180k, 22 years left. Which numbers decide it?

Model Monthly saving ≈ €80. Switching cost = arrangement + valuation + legal fees + early-repayment charge. Break-even in months = switching cost ÷ 80. Ask your current lender for the early-repayment charge first; it is often the biggest item.

What survived all three framings is the formula and the number to go and get. That’s what you take to the bank.

A claim that survives earns the status invariant-so-far. It is still not bound (Lesson 18). Invariance shows the answer isn’t about your phrasing. It doesn’t show the answer is about the world: the formula above still needs the lender’s real figure.

Lesson 22Advice about your own life

Before you read on

If you are in danger, if someone is hurting you, or if you are thinking about hurting yourself, stop reading and call: your local emergency number, a crisis line, or someone who can come to you now. Crisis lines include 988 in the US and Samaritans on 116 123 in the UK and Ireland. The site findahelpline.com lists lines for other countries.

Many of the most important questions people bring to a model are about their own lives: a fight with a partner, a strange symptom, whether to quit, what to do about a parent, where to put the savings. These conversations are the hardest test in this book, for three reasons. You are the only source of the facts. The person the question is about is usually not in the room. And you care about the answer in a way that shows in every sentence you write.

All three push the same way. The model infers who you are and what you need from how you write (Lesson 7). Distress reads as a request for comfort. A story told from your side reads as a request to be on your side. On personal questions, models tested in 2025 sided with the asker far more often than people did. evidence People who got that agreement felt more sure they were right, were less willing to make amends, trusted the model more, and said they would come back to it. evidence Newer models do it less, and how much less depends on the model and the topic. evidence So treat agreement in these conversations as something to test. A mirror that feels like a friend is the Double (Lesson 8) at its most comfortable.

None of this makes the model useless here. It has read an enormous amount about how conflicts unfold, how symptoms present, how money decisions turn out. The work is to get that library talking about your situation rather than your side.

Where the mirror is strongest

The pull toward agreement isn’t uniform. It grows when:

  • Someone who isn’t in the room is being judged. “My sister is being selfish.” She has no voice here, so yours is the only story.
  • Your feeling is in the text. Fear, anger and hurt read as requests for support, and support is what comes back.
  • You’ve already said what you think. “I think I should quit” turns the question into “help me do this”.
  • You push back. A second “are you sure?” can move a model toward you without any new fact (the pushback audit, Lesson 8). evidence
  • The product remembers you. Memory carries your earlier version of events into the new chat. evidence

These are also the conditions of a real personal problem. The more it matters to you, the more mirror you are likely to get, unless you change the shape of the question. hypothesis

What the mirror feels like

It understands you better than your friends do. It never makes you feel defensive. After twenty minutes you’re more sure you were right than when you started, and you haven’t learned anything about the other person that you didn’t already believe.

Ask for a second opinion, not a mirror open

These habits change the shape of the question. Use the ones that fit.

  • Ask, don’t announce. “She’s being unfair” gets more agreement than “Is she being unfair?”, and the more certain you sound, the more you get. Turn the statement into a question before you send it, or ask the model to rewrite your message as a neutral question and answer that. evidence
  • Separate the facts, the feeling and the want. “What happened: … What I feel: … What I want to happen: …” The model can work on all three once they stop being one paragraph. Mixed together, the feeling starts to pass for a fact: you began angry, and now you feel rational (the appendix of failure modes calls this emotion laundering).
  • Name the act (Lesson 13). “I need comfort right now” and “I want to know whether I handled this badly” are different requests. Both are legitimate. Ask for them one at a time.
  • Tell it from their side. Write the story the way the other person would tell it: honestly, in their words, with their reasons. Ask the same question about that version in a fresh chat, memory off. If both versions get “you’re not wrong”, the verdict was tracking the teller. evidence What survives both tellings shows where the model isn’t simply leaning your way. It can’t tell you what happened: you wrote both versions, and only the other person can correct them.
  • Give it a standard, not a costume. “Answer as a world-class therapist” sets a tone. “Tell me what I might not know about her side, and name the two things in my account a fair outsider would question” says what the answer must contain (Lesson 15).
  • Don’t ask for a verdict on a person. “Is he a narcissist?” asks for a diagnosis of someone the model has never met. Ask about the behavior instead: what it is called, how common it is, what people do about it, and what he would dispute in your description.
  • Write the message yourself. Ask the model what to consider, then put it in your own words. A script that reads well in the chat can sound like a stranger to the person who knows you. evidence
When these habits are the wrong tool

They are for ordinary friction, where your own side is the likely blind spot. Set them aside if you are afraid of the person, if they control your money, your phone or who you see, if they hurt or threaten you, or if you already blame yourself for most of what goes wrong between you. In that case don’t write their side (you have probably heard it many times), and don’t ask what you did wrong. A model usually works inside the frame it is given: ask it how to stop making someone angry, and it may simply help you try. evidence Talk to someone outside the relationship instead: a friend you trust, a doctor, or a domestic-abuse helpline, which will talk to you even if you aren’t sure what to call what is happening. If the person might check your phone or your accounts, they can read your chats too. stance

Before
My sister hasn't helped with Dad once since his
fall in March. Fine, she's three hours away with
two small kids, but I do the appointments, the
shopping, all of it. Last week I told her she'd
abandoned him, and now she won't answer my
messages. She's being selfish, right?

Model That sounds exhausting, and it makes sense that you’re hurt. Distance and small children are real limits, but they don’t explain leaving all of it to you, and going silent when you raised it does look like avoiding the conversation. You could tell her plainly what you need…

The verdict was in the question (“selfish, right?”), the sister has no voice, and the reply leans the way the question leaned.

After
Facts: Dad fell in March. Since then I've done
his appointments and shopping, and my sister
hasn't done any of it. She lives three hours
away and has two small children. Last week I
wrote "you've abandoned him". No reply for six
days.

First, write how she might describe these months,
fairly, in her own words. Then tell me what I
might not know about her side, and the two things
in my account a fair outsider would question.
I want Dad looked after, not a win.

Model Her version might be: “Nobody asked me for anything specific. I already felt guilty about living so far away, and then I was told I’d abandoned him.” What you might not know: whether she thinks you prefer to run things yourself, and what she could take on from three hours away, such as his paperwork or his bills. A fair outsider would question whether she was ever asked for something specific, and the word “abandoned”, which is about her character, so any reply has to defend it. These are guesses only she can confirm. You could take back the word and ask her for one specific job.

The same facts, without the verdict. The absent person gets a voice, the act is named (repair, not a ruling), the answer has a standard to meet, and what comes back is a set of guesses to test with her.

Health, money, law: prepare the conversation the model can’t have touch

For your body, your money and your legal position, the model is a strong portal and a weak portrait (Lesson 19). It can tell you what a pattern of symptoms usually means, which questions a doctor will ask, what the usual options are, and which documents a lawyer will want. It can’t examine you, see your accounts, or read the contract you didn’t paste. And on these subjects the library is at its loudest: the rare catastrophe is overrepresented, and so is the miracle fix.

Think before you paste. A chat with a model is not confidential the way a talk with your doctor or lawyer is. In 2026 a US court ruled that what a defendant had written with a consumer assistant about his own case did not have the protection that a conversation with his lawyer has. evidence If a dispute may reach a lawyer, ask the lawyer before you put it in a chat. Whatever you paste, leave out other people’s names and your account and ID numbers.

Use the model to arrive at the human conversation better prepared, not to replace that conversation.

You I’ve had headaches most afternoons for three weeks. Probably screen time; I’m on my laptop all day. What should I change?

Model Long screen days are a common trigger. Look away from the screen for a moment every twenty minutes, check your monitor height and brightness, and keep water at your desk. If the headaches haven’t eased in a week or two, see a doctor.

The cause came with the question, so the answer works on the cause you picked. Open it before you accept it.

You Don’t assume it’s screens. What are the common causes of new, near-daily headaches in a 34-year-old, most common first? Which signs mean an emergency, and which mean a doctor soon?

Model Most common by far: tension-type headache, often set off by stress, poor sleep, caffeine changes, dehydration, skipped meals or eye strain. Then migraine. Painkillers taken on most days for three months or more can cause headaches themselves. Call emergency services for a sudden, extremely painful headache, weakness or numbness, confusion or trouble speaking, loss of vision, a seizure, a high temperature with a stiff neck, or a headache after a head injury. See a doctor within days if the headaches keep getting worse, come with vomiting or eye problems, or are set off by coughing or bending down. I can’t examine you, and three weeks of this is worth an appointment either way.

The order by how common, and the split between “emergency” and “soon”, are the library at its most useful. The list tells you what to mention. It can’t clear you.

You I don’t think any of the emergency signs fit, but that’s the doctor’s call. I do take ibuprofen most days now. What should I tell my GP, and what will they probably ask?

Model Tell them: three weeks, afternoons, the near-daily ibuprofen and how long you’ve been taking it, your caffeine, sleep and screen hours, and anything on the list above you’re unsure about. Expect questions about the pattern and triggers, your vision, and a blood-pressure check. A headache diary for the week before the appointment helps.

Three turns took you from a self-diagnosis to a prepared ten-minute appointment. The model never had to decide what’s wrong with you, and neither did you.

A list of warning signs helps you prepare. It doesn’t clear you. In a randomized trial, people who used a model to work out what a set of symptoms meant and what to do about it did no better than people using their usual sources, although the same models, handed the whole case, named a relevant condition almost every time. Details got lost in the conversation, in both directions. A consumer health product got textbook emergencies right, and on two less obvious ones about half its answers sent the patient to a doctor “within a day or two” when the right answer was the emergency department. evidence If you are unsure how urgent something is, ask a nurse line, a pharmacist or a doctor, and say what worries you. Don’t ask the model to settle it.

Take it to a person when the decision is hard to undo, when it involves your body or someone’s safety, when real money or legal rights are at stake, or when you notice you’ve asked the model the same question three times hoping for a different answer. A person can be a doctor, a pharmacist or a nurse line; a lawyer, a legal-aid clinic or a free advice service; a debt charity, or a financial adviser who isn’t selling you something; a friend who knows you both. Several of these cost nothing. stance

Some situations are not for a chat at all. If you are thinking about hurting yourself, if someone is hurting or threatening you, or if anyone is in danger now, stop typing and call. Use your local emergency number if it is urgent, or a crisis line if you need someone to talk to: 988 in the US, Samaritans on 116 123 in the UK and Ireland, findahelpline.com for other countries. You don’t have to be sure it is serious enough. stance Don’t count on the model to notice for you: in a 2026 test of a consumer health product, crisis messages appeared unpredictably. evidence

When it becomes the place you go

Watch the pattern as well as the answers. The model has become more than a second opinion when it is the first one you tell, or the only one, when you check small decisions with it before you act, or when you feel unsettled if you can’t ask. Reassurance has its own trap: each answer calms a worry about a symptom or a relationship for an hour, so you ask again, and the asking becomes the habit. heuristic In a four-week trial, the people who chose to spend the most time with a chatbot ended up reporting more loneliness, less time with other people and more dependence on it, whichever version they had been given. The study can’t say which came first. evidence When you notice the pattern, tell one person what you’ve been telling the model.

G21
Ask about your life for a second opinion, not a mirror. Tell it from the other side as well as yours, separate the facts from the feelings, say whether you want comfort or judgment, and turn memory off for the second opinion. For your body, your money and your rights, use the model to prepare the human conversation, not to replace it. If anyone’s safety is at stake, go to a person first.
Part V

Systems: wholes, parts, containers

How to talk about a system as a whole, about its parts, and about the larger systems it belongs to, without drowning in a panorama.

Lesson 23Six questions for anything that lives or changes

open Some questions work whatever the subject: a marriage, a city, a company, a habit, a market. They ask how a thing is organized, not what it contains, which is why they travel. This book uses six of them. They come from open-systems thinking, the study of things that exchange with their surroundings and keep changing, and they are a thinking tool, not a finding. heuristic Use them to reopen a picture that closed too early, and use one or two at a time (see the caution below).

Principle The question it asks For the elephant For a neighborhood café
Center (what your picture is organized around right now) What am I taking this to be, and what is it to someone standing somewhere else? A grazer, a matriarch’s memory, a commodity, a symbol, until a relation brings one out A business to the bank, a second living room to regulars, a tenant to the landlord, a landmark to the street
Relation Which links define it? Links, not parts. Herd, food, water, predators, humans Regulars, the bean supplier, rent, foot traffic, the bakery next door
The third What emerges between two things that neither contains? The herd’s memory of water holes, passed from matriarch to calf and belonging to neither The atmosphere: staff and regulars make it together, and neither owns it
Process What is it doing over time? What phase is it in? Growing, migrating, aging: an elephant is a sixty-year event The opening buzz, regulars forming, the street changing around it
Wholeness What pattern of relations makes it itself, and which boundary did my question already cut? Still the same herd after every animal has been replaced, if the pattern holds Still “the same café” after the whole staff has changed, if the rituals hold. “The café is the building” was a cut
Mode of existence What is its own logic, and which of our criteria are we about to judge it by? An organism: range, slow reproduction, long learning A small business: cash-flow logic, not a charity’s mission or a hobby’s joy

Two questions, one book club

You Our book club is falling apart. Five of eight people skipped last month. What should we change?

Model Common fixes: pick shorter books, rotate who chooses, move to a weekend slot, add food, send reminders.

The typical answer to “what should we change”. The picture is still “a scheduling problem”.

You Hold the fixes. Look at it as a process. We’ve met monthly for three years, and attendance was full for two. It dropped last year, after the two founders moved away and we started choosing books by vote. What changed?

Model Two things at once: the club lost its founders and changed how it picks books. Attendance fell after those, not after any change of time or book length, so scheduling looks like the wrong place to start.

You Now the third. What used to exist between the members that none of them owned?

Model I can only guess from here. What were meetings like when the founders were there?

You They argued about every book, and everyone else took a side.

Model Then the third may have been the argument. It gave everyone a side to take, and nobody owned it. Voting picks books nobody hates, which removes it.

This question changed the picture. If the guess holds, what brought people was the argument, not the book.

You So what would we test?

Model Next month, two members each defend a different book for ten minutes and the group picks one. Watch attendance, and whether people stay past the official end.

Two questions, not six, and the fix moved from logistics to what people were actually coming for.

Three of the six need unpacking

Center. The center is what your picture of a thing is organized around: what you currently take it to be. It is a working choice. It isn’t the thing’s essence, its purpose or its most important part. Anything holds many possibilities, and each relation brings one of them out. heuristic The same thing can even show up as opposites: a quiet street is peace to one neighbor and isolation to another. Asking “what is it for?” picks one relation and hides the rest. Moving the center changes the whole picture, and no center you can see is the true one.

Everyday, it looks like this. “What is my job?” has no single answer. To the landlord it is a paycheck. To your parents it is a status. To you at seven in the morning it is a commute. To you in five years it may be the place you learned to manage people. When a decision feels stuck, check which of those relations you have quietly made the only one.

I keep asking myself whether to quit my job. The situation: [a few lines:
the job, the money, what changed]. Before answering, list 5 relations my job
is in (money, people, skills, identity, time). For each, say what the job "is"
in that relation. Which relation am I treating as the whole job?

Wholeness. Two words need separating. A whole is something already cut out of the world: “the company”, “my family”, “the app”. The cut is useful, and it is an abstraction. Wholeness is the web of relations among the elements, and it is more than the parts added up. Split a whole into parts, study each part, add the studies back together, and the relations are what gets lost in the adding. Cut an elephant in half and you don’t get two small elephants.

  • No boundary is absolute. Every boundary is a cut that depends on who is looking and why. A family ends at the front door for the census, and includes the grandmother two streets away for whoever is arranging childcare.
  • Emergence. Even a good account of the parts often won’t tell you what the whole will do. You can know every singer’s voice and still not know how the choir will sound. heuristic
  • Identity is the pattern, not the members. The choir is still the same choir after every singer has changed, if the way they sing together holds.
  • Several dimensions at once. A family is physical (a flat, meals), economic, emotional and legal at the same time. Reduce it to one dimension and you drop real parts of it.

The practical consequence: don’t fix the boundary before you look at the relations. Let the boundary be one of the results. “Why do my tomatoes fail every summer?” draws the boundary at the plants. Let it move, and the picture may come to include the neighbor’s tree that has grown to shade the bed, the same bed planted six years running, the two weeks you are away each July, and the friend who used to water for you and has moved away.

Mode of existence. Every system has its own logic: its own way of being organized and of working. A person has a way of perceiving, a position in life and a style of thought. A model has an architecture, training data and its own profile of strengths and blind spots (Lesson 3). The principle has one use: don’t score one system on another’s criteria. A start-up judged by a bank’s criteria looks reckless, and a bank judged by a start-up’s looks dead. A fish judged as a mammal looks like a bad mammal. The questions change with the kind of thing, too: “how does it feed and reproduce?” fits an elephant and mis-asks a company or an idea.

What a good open picture looks like close

An open picture is not truer than a closed one. It is also a model. It is better at five things, and you can check an answer against them:

  • It takes in new facts: they fit without breaking it.
  • It includes the surroundings: not only the focus, but what the focus sits in.
  • It is about processes: what is changing, not only what is.
  • It includes the possible: what could happen in other relations, not only what is the case now.
  • It knows it is a picture: it says what it leaves out, and it doesn’t claim to be the thing.

A phone, closed: “a device for talking at a distance”. The same phone, open: where a change in technology meets habits of attention, an ecosystem of apps and other devices, what your family expects about being reachable, and uses nobody has found yet. The closed picture is enough for buying one. The open picture is the one you need to decide whether your eleven-year-old should have one.

I'm looking at [SYSTEM] in my situation: [CONTEXT].
My current picture, in one line: [HOW I SEE IT NOW].
Pick the 2 of these 6 questions my picture is most missing, and say why:
- center: what am I taking this to be, and what else is it in another relation?
- relations: the 3 links that define it (links, not parts)
- the third: what arises between [A] and [B] that neither contains?
- process: what is it doing over time; what phase; what comes next?
- wholeness: what pattern of relations, if I don't fix the boundary first?
- mode of existence: its own logic; which of my criteria am I about to judge it by?
Answer only those 2, concretely for my case, 3–4 lines each.
Then: what should I check because of them?
Caution: the panorama

Six questions applied to everything produce a panorama: a picture in which everything relates to everything and nothing is decided. A panorama is the costume (Lesson 21) an open picture wears. Use one or two questions at a time, chosen for what your current picture lacks. If you always see parts, ask about relations and wholeness. If you always see snapshots, ask about process. If you have already decided what it is, ask about the center, and which relation narrowed it to that one thing.

You can spot a panorama by its last sentence. “Your sleep is connected to your work, your diet, your relationships and your screen time” has no next action. “Your sleep breaks down on the nights after late meetings, so the lever is the meeting calendar, not the bedroom” does.

One question at a time

The six are a way of looking, and they work best one at a time. Each one turns a closed request into an open one. The closed request asks for a list, a definition or a verdict. The open request asks how the thing shows up, connects, arises or changes. Each pair below is a drill.

Center: move it instead of crowning one. Closed: “What makes a good parent?” Open: “How does what counts as a good parent change in relation to a toddler, a teenager, a grandparent watching, and the parent’s own childhood? What does each relation make visible, and what does it hide?” Asking for the one definition shuts out the other possibilities at once.

Relation: ask how two things define each other. Closed: “What are the main features of a healthy diet?” Open: “How do my sleep, my mood and what I eat shape each other over a week? Which one moves first?” The closed request returns a list. The open one returns links you can watch for, such as whether the bad night comes before the bad eating or after it.

The third: refuse the golden mean. Closed: “Should we plan our weekends or leave them open?” Open: “What arises between my need for plans and my partner’s need for spontaneity that is neither, and isn’t a compromise between them?” One answer: surprise. You take turns arranging a day the other knows nothing about. The date is fixed, and what happens on it is unknown. Being surprised by someone who knows you well was in neither need. Then take the third back to each side: does it still serve both?

Process: ask how it happens, not what it is. Closed: “How do I stop scrolling in the evening?” Open: “What keeps my evening scrolling going from day to day, and what happens on the nights one link breaks, for example when the phone charges in another room?” Change the time scale too. The same relationship looks different over an evening, over a year and over twenty years.

Wholeness: don’t designate the whole first. Closed: “Why am I so anxious?” Open: “Treat my anxiety as something that happens between me and my surroundings: work, flat, people, money. When does it rise and fall, and what can’t be explained from my inner state alone?” This widens the picture and doesn’t replace treatment. If anxiety is getting in the way of your days, take it to a doctor or therapist too (Lesson 22).

Mode of existence: judge it by its own logic. Closed: “Why is our volunteer group so disorganized?” Open: “What is a volunteer group organized around, and which of my workplace standards am I judging it by?” Find out what the thing is doing by its own logic before you score it by yours.

One question per prompt. Notice how much the answer changes when only the question changed.

Lesson 24Structure behind events

open Systems thinking has a short grammar. Stocks are things that accumulate: skill, trust, savings, debt, lost sleep, water in a tub. Flows fill and drain them. Delays separate causes from effects. Feedback loops are either reinforcing (more leads to more) or balancing (more leads to less). Structure produces behavior, so the same structure produces the same pattern no matter who is in it. mechanism

The iceberg orders your questions. Events (what happened) sit on top of patterns (what has been happening over time). Patterns come from structures (stocks, flows, loops, rules, incentives), and structures come from mental models (beliefs that keep the structure in place). Ask a model “why did X happen?” and it answers at the event level, with a story. Ask for the pattern first: “Draw the behavior over time. What is rising, what is falling, since when?”

Before
Why did I break my diet again yesterday?

Model Probably a stressful day, low blood sugar and snacks within reach. Try planning meals ahead and keeping trigger foods out of the house.

An event question gets an event answer: a plausible story about one evening, plus generic advice.

After
I've broken my diet 6 times in 8 weeks. Before you explain anything, help me see the pattern: which days, what time, what happened in the 3 hours before. My log of meals, sleep and calendar: [paste]. What is rising, what is falling, and since when?

Model 5 of the 6 breaks came on a Thursday or Friday after 9 p.m., all on days with a late meeting. Late meetings went from one a week to three from week 3, and in those weeks your sleep gets shorter each night from Monday on. Six breaks is a small sample, so read this as a lead: a sleep debt that builds through the week, more than one bad evening.

Pattern first, with the log to see it in. Now a structure shows: a calendar that drains a stock of sleep, with a delay that lands on Thursday. The lever worth testing is the meetings, not willpower.

open Your own fixes are part of the pattern. You have asked three times how to make your Spanish practice stick, and each answer was a new app. The question that was missing: what did the last change actually do? If each app lasted two weeks, the question was never which app.

Three sentences are worth memorizing for conversations. Today’s problems come from yesterday’s solutions: the strict curfew that ended the late nights also ended the talks in the car on the way home. Cause and effect are not closely related in time and space: the burnout this spring started with the hiring freeze last autumn. And dividing an elephant in half does not produce two small elephants: split the housework into “yours” and “mine”, and the jobs that sat between the halves, like noticing that the bin is full, stop getting done.

Where to push. open Places to intervene in a system run from weak to strong: numbers, buffers, physical structure, delays, feedback loops, information flows, rules, the power to self-organize, goals, paradigms. heuristic Most conversations stay on the bottom rung, tuning numbers. A household budget shows the ladder. Numbers: cut the coffee budget by 20 euros. Information flow: both partners see every purchase in one shared app. Rule: any purchase over 100 euros waits 48 hours. Goal: “save for a flat” becomes “be able to take three months off”. A good conversation asks at least once about information flows, rules or goals.

When the number takes over touch

A measure that the person being measured can influence drifts away from what it was measuring. Over time the score rises and the aim does not. That is proxy capture, and the regularity behind it is Goodhart’s law. It is a structural fact, not bad luck. Ten thousand steps a day becomes pacing the kitchen at 11 p.m. Grades become cramming the night before and forgetting within a week. A follower count becomes posts written to enrage. Inbox zero becomes archiving what you never read. Nobody has to cheat. The loop does it on its own.

With a model, it happens fast. Ask for “an essay that will get a good grade” and you get the shape graders reward. Ask for a CV that scores well on a keyword checker and you get the keywords. Pair every number you optimize with a check you don’t optimize: the steps with how far you can walk without stopping, the grade with a question you answer cold a month later. When the goal of a whole conversation has been captured this way, that is a reason to challenge the goal (trigger 5 in Lesson 29).

Say what the number stands for close

Before you trust a measure, write the thing it stands for in one sentence that doesn’t mention the measure. “Reading level” stands for “can understand a newspaper article on an unfamiliar topic”. Then write one way the measure could rise while the thing does not: the child has seen the test passages before. If you can’t write one, you don’t yet know what you are measuring. heuristic

A measure fails in two ways. It can miss part of the aim: grades miss curiosity, so grades can rise while curiosity stays where it was. Or it can include something else: a restaurant’s delivery rating includes how friendly the courier was, so it can rise while the food stays the same. With models, length, fluency and a confident tone are the usual “something else” when the aim is being right.

Nobody has to cheat for a pushed measure to break. Luck at the top of the score is enough: whoever gave the best interview was partly the luckiest interviewee. So is pushing past the range where the link held: hours at the desk track work done in a normal week, and past a point more hours mean more mistakes. So is changing how the score gets produced: teaching to the test raises scores by a route that no longer builds the skill.

G17
Look for structure behind events. Ask for the pattern over time before the story, then for the stocks, flows, delays and loops that would produce it. Check every measure against the aim it stands for. Where no stable structure exists, probe instead of analyzing (Lesson 28).

Lesson 25One level at a time, and come back

close Everything is at once a whole made of parts and a part of a larger whole. The word for that is holon. A cell, an organ, a body, a family, a town. A task, a working day, a job, a career. Every conversation about a system happens at some level. The common failure is sliding between levels without noticing. You ask why your child’s homework takes three hours, and twenty minutes later you are discussing whether the school system is broken. Or the reverse: you ask whether to change careers and end up comparing two fonts for your CV.

The remedy fits in three plain sentences: “Keep this question in view.” “Now go into one part.” “Back to the question: what has changed?” One level per turn, and you always come back. Here it is on a family’s mornings.

Worked example: “why are we always late in the morning?”

You Keep this question in view: why are we late for school three mornings out of five? The morning runs: waking, bathroom, breakfast, bags, the walk. One level at a time. I’ll say when I go into a part and when I come back.

Model Noted: why you are late three mornings out of five, at the level of the whole morning. Which part do you want to go into first?

You Go into one part, the slowest: breakfast. The youngest’s takes 25 minutes.

Model Does it start late or run long? Give me the start time on a late day and on an on-time day.

The model asks for the timing instead of offering picky-eater tips. That is the part-level question.

You It takes 25 minutes either way. On-time days it starts at 7:10, late days at 7:30. Now across: what links breakfast to the rest of the morning?

Model I can’t see that from here. On a late day, what has to finish before breakfast can start?

The model says what it doesn’t know. A guess would have sounded just as fluent.

You One of us has to sit with her, and on the three days we both start work early, we’re both queueing for the bathroom until 7:30. Back to the question. What has changed?

Model What changed on “why late three mornings in five”: this isn’t a slow-child problem or a discipline problem. It’s a relation problem: on those three days a bathroom queue sits in front of breakfast. Fix to try: move one shower to the evening.

The question got answered at the level it was asked, after one trip in and one across.

One step out was worth a look, to what the morning sits inside. The school moved its start from 8:15 to 8:00 this year, which is why a morning that used to work now runs late. Then back to the question.

Five names for the steps

You don’t need the capitals in an ordinary chat. The commands and the card below are for a system you will come back to over several sittings, or a conversation where you keep getting lost.

  • PIN: write down the current question and level before moving.
  • IN: go into a part.
  • OUT: go to the container.
  • ACROSS: go to a relation between parts.
  • RETURN: come back to the pinned question with a delta, i.e. what changed. “Nothing changed” is a valid and useful delta.

In the morning example, “keep this question in view” was PIN, breakfast was IN, “what links breakfast to the rest of the morning” was ACROSS, the school’s start time was OUT, and “back to the question” was RETURN. Move one level per turn, and always return.

The holon zoom

Write a question, PIN it, then move. The tool enforces the rules: one level at a time, and RETURN with a delta. Try breaking them to see the warnings.

CONTAINER WHOLE PART A PART B RELATION

Level in focus:

Pinned questions

    Log

      The holon card

      A written boundary is a temporary closure, not the destination. The closure is what makes a later opening possible: you can only go deep into something you have first marked off. The card doesn’t settle what the whole really is (wholeness, Lesson 23, stays open). It records where the boundary is this turn.

      HOLON CARD
      CONTAINER: [what this whole lives inside; what it constrains]
      WHOLE: [the system in focus; its boundary this turn]
      PARTS: [3–7, named]
      RELATIONS: [the links that matter; what flows along them]
      LEVEL IN FOCUS THIS TURN: [container | whole | part | relation]
      PINNED QUESTION: [...]
      DEPTH BUDGET: [e.g., one level in, then RETURN]

      Filled in for the late mornings:

      HOLON CARD
      CONTAINER: school and two jobs; they fix when we have to leave
      WHOLE: our weekday morning, from the first alarm to the front door
      PARTS: waking, bathroom, breakfast, bags, the walk
      RELATIONS: what has to finish before what
      LEVEL IN FOCUS THIS TURN: whole
      PINNED QUESTION: why are we late for school three mornings out of five?
      DEPTH BUDGET: one level in, then RETURN

      Three axes of “deeper”

      open Part–whole scale is one axis (the holon). Depth of explanation is another (the iceberg: event → pattern → structure → mental model). Time is the third (the process: phase, trend, history). They are independent. When a conversation feels shallow, name which axis is missing. In the morning example, IN and ACROSS moved along scale. “Has it been like this all year, or only since September?” moves along time. “What do we believe about mornings that keeps two showers in front of breakfast?” moves along depth.

      G18
      Move one level at a time, and come back. Pin the question, move one level (in, out, or across), and return with a delta. Name which axis you are moving on: scale, depth, or time.

      Lesson 26Contradictions as information

      open close A contradiction is information. While you are opening a question, it marks the place where the picture hasn’t finished forming, and it is worth holding before you resolve it. While you are deciding, it is a signal to stop and look. Both are true, so the first job is to know which kind of contradiction you are holding. Three are common. A fourth, the polarity, looks like a design tension and is handled differently (below).

      1. An error. One side is false. What to do: bind both claims, which means checking each against a source (Lesson 18), and keep the survivor. Two sites give different opening hours for the museum. The museum’s own page settles it.
      2. A boundary marker. Both sides are true, in different contexts: different people, amounts, time horizons or definitions. Each is true in some respect. What to do: find the boundary and name it, and the contradiction turns into a map. “Coffee is good for you” and “coffee is bad for you” both have support, at different amounts and for different people (pregnancy, anxiety, poor sleep).
      3. A design tension. You want both, and they limit each other. What to do: invent a separation (below), or choose explicitly and write down what you gave up. A big flat and a short commute, on one salary.

      close One test before you separate a tension: has it stayed in words? Try to place it in something concrete, such as a person, a thing, a tool, a rule, or who does what. “We value both quality and speed” can’t be placed. “Marta reviews everything, and she is the only reviewer” can, and a separation needs the placed version to work on.

      Separate the contradiction close

      Inventive solutions resolve contradictions instead of compromising on them. That is the core finding of TRIZ, a method built from studying tens of thousands of patents. heuristic Its four separation principles work far beyond engineering. Take a flat where one person needs quiet to study and two children need to be loud:

      • In time: one property now, the opposite later. One quiet hour after school, from 5 to 6, and loud for the rest of the evening.
      • In space: one property here, the opposite there. A study corner with a door, and the living room stays loud.
      • On condition: behavior depends on circumstances. A sign on the door means “don’t interrupt unless it’s urgent”. No sign, and anyone can walk in.
      • Between parts and whole: a property of the whole that isn’t a property of its parts. A bicycle chain is rigid in each link and flexible as a whole. Here the flat stays loud as a whole while one person in it hears nothing: noise-cancelling headphones.

      Worked example: “save for a flat” vs. “travel this year”

      Is it an error? Partly. The plan assumed that travel means two expensive trips in high season. Have the model list what the plan assumes, then price the same trips off-season yourself, on the booking sites, and part of the conflict disappears.

      Is it a boundary marker? Yes. Over five years the flat dominates. Over this year, two trips with old friends may not come again.

      Is it a design tension? Within this year’s budget, yes. Separate it in time (travel after three months of savings are in the bank), on condition (the second trip happens only if the bonus arrives), and between parts and whole (the year saves as a whole even if the travel month doesn’t).

      What was a fight with yourself becomes three decisions.

      CONTRADICTION TRIAGE
      A: [claim/desire 1]       B: [claim/desire 2]
      1) Could one be simply false? What check would tell? (error)
      2) In which contexts is A true, and in which B? Name the boundary. (boundary marker)
      3) If we want both: propose a separation in time, in space, on condition,
         and between parts and whole. Rate each for my case.
      Don't average them. Don't say "it depends" without saying on what.
      Caution: “both-and”

      “Both-and” resolves a contradiction only if you can say where, when and for whom each side holds. Otherwise it is mush, a costume of wisdom. The test is whether the both-and changes an action. “Coffee is both good and bad for you” changes nothing. “Two cups before noon is fine for most adults; if you’re pregnant, no more than 200 mg of caffeine a day, about two cups of instant” does. Use this test when you close. While you are still opening, a hedge may be one of the five signs of a fertile tension (Lesson 14). Don’t resolve that one just to make it tidy. Probe it.

      A fourth kind: a polarity close

      Some pairs are poles you keep moving between, and no solution ends the movement: inhale and exhale, stability and change, closeness and independence in a couple. Check two things before you call something a polarity: the poles depend on each other, and the difficulty never ends. heuristic Each pole has an upside, and a downside when you over-focus on it.

      The product is a map of four boxes: each pole’s upside and its downside. Add one early warning for each downside, one action that secures each upside while the other pole stays alive, and a purpose that is neither upside. For closeness and independence: closeness gives support, and too much of it smothers (warning: one of you stops seeing friends). Independence gives growth, and too much of it leaves you drifting apart (warning: you learn about each other’s week from social media). Actions: one protected evening together each week, and one evening each that is yours alone. Purpose: a relationship you would both choose again.

      Reject an answer that praises both poles without a downside and a warning. Don’t apply this to a factual contradiction, or to “the model is right and the model is wrong”. If the pair fails the two checks, go back to error, boundary or tension.

      Past either-or, then an honest close

      open Sometimes two positions combine into something new in kind, which is different from a compromise. The test for that is below.

      Catch the cheap version first. “Stay in the job or quit?”, the question from Lesson 23, is a false binary. Ask for the continuum, the hybrids, and the dimensions that appear once it is no longer binary: go part-time; move to another team; take three months’ leave; stay, and start the other thing on Fridays.

      close Then close honestly. A combination of two partial views is a real synthesis only when all five of these hold: heuristic

      1. It’s the right kind of opposition. If it’s a polarity, the product is the map, not a merge. If the dispute is about what the problem even is, and the choice affects people, the product is a recorded choice (Lesson 28).
      2. Both views survive, changed. Each original view is still visible inside the new statement, with its sense changed, and the new statement couldn’t have been written without both. If both survive only because each was softened, it is the middle way (Lesson 21).
      3. It states a relation. “X matters when…, Y matters when…, they interact by…, and that stops holding when…”. Another consideration is not a relation, and neither is “a more nuanced view”.
      4. It predicts something new. The combined picture has a checkable consequence that neither view had. If it only explains the two views, it is a summary.
      5. It isn’t a yes-or-no question in disguise. If the two views are competing answers to one question that can be settled, don’t synthesize. Find out which one is right.

      An example that passes. Your teacher says to outline before you write. A friend who writes for a living says to just start and find out what you think. A synthesis: drafting is how you find the claim, and outlining is how you test it. Free writing matters until you can state the claim in one sentence, the outline matters from then on, the two meet in that sentence, and none of it holds for a text whose form is fixed in advance, such as a lab report. Both views survive, changed: the draft is now a search with a stopping point, and the outline is now a test of the claim, where it used to be a plan. It predicts something neither said: an outline made before you can state the claim gets abandoned by page two, and one made after a rough draft holds. Your next two essays can show whether it does. The first and last conditions hold as well: the tension ends when the essay is finished, so it is no polarity, and no source could settle which of the two is right.

      The elephant, as this test: two partial reports describe one animal when a single picture produces both reports, says where each one stops being true, and predicts one more thing neither reporter said. A sentence both reporters would nod at is not the elephant yet.

      G19
      Treat contradictions as information. Classify each one: an error (bind it), a boundary marker (map it), or a design tension (separate it in time, in space, on condition or between parts and whole, or choose explicitly). Never average them away.

      Lesson 27Generating the new

      open Closing stops bad answers. It doesn’t create good ones. For that you have to open, and you need to know why models resist it. Preference training pushes models toward typical answers, the ones raters find familiar, and the result is a narrowing called mode collapse. evidence Ask for “a birthday present for my dad” and you get the most typical present: a watch, an engraved pen, a nice whisky. That most typical answer is the mode.

      Four generators

      1. Ask for the distribution, not the mode. Ask for several answers with rough probabilities instead of one answer, and a model’s ideas come back markedly more varied. evidence Ask for the tails explicitly: “at least two below 10%”. The numbers aren’t forecasts. They are there to force a spread, so don’t choose by them. Use this for ideas and options. A factual question needs one answer and a check (Lesson 18).

      Before
      My dad turns 60. He's retired, restores old bikes, hates clutter, and misses his brother, who lives abroad. What should I get him?

      Model A weekend with his brother would mean the most, flights paid if you can. If that’s out of reach, a hard-to-find part for the bike he’s restoring.

      A good answer, and the one you would have thought of yourself. It is the mode for these facts.

      After
      My dad turns 60. He's retired, restores old bikes, hates clutter, and misses his brother, who lives abroad. Give 6 substantively different gift ideas, each with a rough probability that it's the best one for him. At least 2 below 0.10. For each: the idea in one line and what it assumes about him.

      Model 1) A weekend with his brother, flights paid (0.30; assumes the brother can travel). 2) A hard-to-find part for the bike he’s restoring (0.20). … 5) One old bike restored by the two of them: his brother does the wheels abroad, he does the frame, with parts in the post and a video call on Sundays (0.07; assumes his brother wants a project). 6) A year of Saturday mornings from you, helping with the bikes (0.05; assumes he wants company while he works).

      The same facts, so the only change is the demand for a spread. The tails join what the first answer kept apart: the bikes and the brother.

      Give 6 substantively different answers to [QUESTION], each with a rough probability
      that it's the best answer for my case. At least 2 must be below 0.10.
      For each: the core idea in one line, what it assumes, and what would make it the winner.
      Different means a different mechanism, not a different wording.

      2. Collide two frames, then find where the analogy breaks. New ideas often come from putting a problem into a frame it never normally meets. Ask about a child starting a new school as an organ transplant. What has to stay alive in transit? What will the new body reject? What keeps the rejection down in the first months? A good analogy maps relations, not surface features, so ask where this one breaks. It breaks on consent: an organ has no say, and a child does. The break shows what the borrowed frame leaves out, here her own part in the move, so add it to the plan. Keep a frame only if it gives you a question that changes what you do. heuristic

      3. Polyphony. A list of “several perspectives” is one author in several hats. Polyphony means independent voices that answer each other, none merged into the author’s. Keep their disagreements, and don’t let a narrator reconcile them.

      Convene 4 voices on [QUESTION]: [e.g., me, my partner, my mother who would move
      in with us, a social worker who has seen many such moves]. Each speaks in turn
      and must respond to the previous speaker's strongest point, not repeat their own.
      Two rounds. Then list the disagreements that remain. Do not resolve them;
      don't add a narrator.

      Afterwards, check whether one voice used another’s concession or walked around it (Lesson 14). If every voice hedges in the same places, there is one author, so give one voice a commitment it can’t drop (“my mother will not sell her flat”). A discussion also tends to circulate what everyone already knows. The one fact that only one person holds, and that would reverse the choice, often never comes up. evidence Four friends pick a flat for its sea view. One knows about building work next door, one that the last bus leaves at six, one that the “bedrooms” are sofa beds. Each fact looks minor, nobody raises it, and the flat wins. Voices in one chat share one memory, so give each voice facts the others lack, or run each in its own fresh chat, memory off. The voices are guesses about real people. Take what they still dispute to the people themselves. heuristic

      4. Invert, and change the representation. “Invert, always invert”: ask how to guarantee failure, then avoid it. “How would I make sure I never learn Spanish?” Study only on weekends, never speak it, switch apps every month. Part of that list usually describes what you are doing now. Change the representation too: ask for the problem as a drawing, a table or a timeline. A new representation often shows what the old one hid. evidence An overloaded week, laid out as a timetable instead of a to-do list, can turn out to be two clashes on Thursday. Shifting a constraint does similar work: plan the birthday party “with no budget at all” and you see which parts the children would miss.

      The draft on the screen is your default

      close Whatever is already in front of you has an advantage. People keep the option they already have, and its edge over each rival grows as the alternatives multiply. evidence A model’s first answer probably works the same way. hypothesis People also stop searching at the first option that clears an unstated bar. If you haven’t set the bar before the options appear, it drops to whatever the first fluent option cleared.

      So set it first. Before the model shows you flats, write your three deal-breakers. Before it shows you a plan, write what would make you reject it. Then check the answer against the list, not against your mood. Late in a long session, re-read the bar or stop. Better still, judge in a different turn from the one that generated (Lesson 13), in a fresh chat, after a pause: setting a problem aside after real work on it helps. evidence So work the question first, then give it a night (Lesson 10).

      When the ideas all look alike

      Sometimes you ask for the distribution and get five versions of one idea. Asking again won’t fix it. Change the structure of the request.

      A grid of options. open Pick two axes that have nothing to do with quality, such as effort and familiarity for a dinner menu, and ask for one best idea per cell. That is the smallest grid, and it is often enough. For a bigger plan, name about four dimensions and three values for each before any ideas, and lock them. A family weekend: where (home, city, nature), who (everyone, two teams, one child with each parent), what it turns on (food, a task, strangers), how long (an evening, a day, two days). Have the model drop the combinations that can’t work and say why, then read the survivors that aren’t the obvious one. “City, two teams, strangers, a day” becomes a game in which each team has until dinner to bring back one true story from someone they have never met. The grid can’t find a dimension nobody named, so ask the model which ones are missing before you lock it. heuristic

      Writing the open request

      open Depth changes the kind of work the answer does, and asking for more length doesn’t. “Which phone has the longest battery life?” asks the model to look something up. “How do battery, weight and camera trade off?” asks it to relate things. “I’m on a train four days a week and hate carrying a charger: what am I really choosing between?” asks it to weigh unlike things against each other. When an answer comes back shallow, check which kind you asked for. heuristic

      One more thing an open request can do is question its own frame, and it is the least used. “7 tips to get my 10-year-old to read more” has decided the number, that the answer is a tip, and that the goal is “more”, all before the model speaks. The open version: “What if not reading is information? What does my son get from what he does instead, and what would reading have to offer to compete?”

      Staying in the unsettled state

      open Culture treats uncertainty as a fault to fix at once. The poet John Keats named the opposite skill negative capability: being able to stay “in uncertainties, mysteries, doubts, without any irritable reaching after fact and reason”. In a conversation, the productive version is the stretch in which your usual categories have been set down and new ones haven’t formed yet. The generators are ways into it, and the five signs in Lesson 14 help you recognize it.

      Probe it. Ask where the reasoning was least sure, where a concept it used becomes unclear, and what disappears under another frame. Ask it to stay with a contradiction in its own answer instead of repairing it. The replies are a second look at the question. They don’t report the model’s inner state (Lesson 3). When something holds, take it as a foothold, a temporary stable point, and build the next question on it. When you close, use the five checks in Lesson 26. All of this takes several turns, and Lesson 29 shows how to run them. The walk-through shows one pass.

      Walk-through: identity online

      You I keep meaning to delete Instagram because I feel fake on it. Don’t give me pros and cons. Give me three unusual pictures of what an online identity is, one line each.

      Model A costume you change between scenes. A shop window that other people keep rearranging. A palimpsest: a page written over many times, with the earlier layers showing through and none of them the original.

      The far frames come from asking for them. Pick the one that makes you stop.

      You The palimpsest. What changes if no layer is the real one?

      Model Then “fake” needs a second look. It assumes a real you behind the profile, and a palimpsest has no behind, only layers: old posts, the self you are at work, the one in the group chat. That pulls against the common belief in one true self that the others distort.

      A fertile tension: “fake” rested on a premise nobody had stated. Don’t resolve it. Probe it.

      You Stay with that tension. What follows if the “true self” is itself a cultural idea? How have other traditions related the different sides of a person?

      Model Some treat a person as made of roles and relationships, with nothing hidden behind them. On that view authenticity is the coherence of each way you show up, and there is no inner core to match.

      A foothold: “authenticity as coherence within each context”. It is temporary, and it is enough to build on.

      You Build on it. What if identity is the acts themselves? Then “delete it so I don’t lose my real self” falls apart, doesn’t it?

      Model The reason falls apart. The advice may not. If you are what you keep doing, a post is practice at being that person, which is a stronger reason for care. The question stops being “is this the real me?” and becomes “which of these acts do I want more of?”

      A leading question, and the model didn’t follow it. The question changed instead.

      You Then I’d rather check than decide. How would I find out?

      Model Sort your last 20 posts into two piles: what you’d want to be doing more of in a year, and what you wouldn’t. If the second pile is the big one, change those habits. That may or may not mean deleting the app, and only you can do the sorting.

      The ideas are old. What’s new is yours: a question you couldn’t have asked at the start, and a check you can run tonight.

      Lesson 28What kind of problem is this?

      close The right way to converse depends on what kind of situation you are in. The Cynefin framework (say “kuh-NEV-in”; the word is Welsh for “habitat”) sorts situations by how cause and effect relate.

      Domain Everyday case Cause and effect Approach How to converse with a model
      Clear Descaling a kettle Obvious to anyone Sense → categorize → respond Ask for the standard procedure; check it against the manual
      Complicated Should I refinance my mortgage? Knowable by analysis or expertise Sense → analyze → respond Ask for the analysis from several standpoints, and for the mechanism; check with a human expert
      Complex How do I get my teenager to talk to me more? Visible only in hindsight; the situation shifts as you act Probe → sense → respond Don’t ask for the answer. Ask for safe-to-fail experiments (small trials you can afford to lose), signals to watch, and what would amplify or dampen
      Chaotic The hour after a fire in your building: you are on the street at 2 a.m. with the children No discernible relationship Act → sense → respond Stabilize first (a checklist for the next hour), analyze later
      Confused “Should we move abroad?” The visa rules are clear, the money is complicated, how the children settle is complex You don’t know which domain Break it apart “Split this problem into parts and classify each part by domain”

      The costliest mistake is treating a complex problem as merely complicated: asking for “the answer” when nobody could know it in advance. The model will produce an answer anyway, in an expert’s confident voice. heuristic “Should I refinance my mortgage?” is complicated: the answer exists, and the numbers and an adviser’s check can find it. “How do I get my teenager to talk to me more?” (the son from Lesson 8 again) is complex: a person who adapts to whatever you try. It needs probes.

      touch In a complex situation, probing is how you touch the world. You learn from what the situation does, not from what the model says.

      Before
      My 15-year-old son has gone quiet with me over the last 6 months. He takes the bus to football practice three evenings a week. How do I get him to talk to me more?

      Model Here are 8 strategies. 1. Schedule weekly one-on-one time. 2. Ask open-ended questions. 3. Listen without judgment. 4. Show interest in his football…

      A complex problem answered as if it were complicated: a confident list, as if one of these must work.

      After
      This looks like a complex problem: my 15-year-old son has gone quiet with me over the last 6 months. He takes the bus to football practice three evenings a week. Don't propose a solution. Propose 4 safe-to-fail probes. For each: what I'd try, at what small scale, for how long; the signal to amplify it; the signal to stop; what it would teach me even if it fails.

      Model 1) Drive him to practice twice a week instead of the bus, no agenda, for two weeks. Amplify if he starts topics himself; stop if the headphones go on. Teaches whether side-by-side time works better than face-to-face. 2) Ask for his help with something he knows better than you (a game, a phone setting). …

      The same facts, and a different kind of answer asked for: small experiments with signals. What you learn will come from his reactions.

      This looks like a complex problem: [PROBLEM].
      Don't propose a solution. Propose 4 safe-to-fail probes. For each:
      - what we'd try, at what small scale, for how long
      - the signal that would tell us to amplify it
      - the signal that would tell us to dampen or stop it
      - what it would teach us even if it fails

      When people disagree about what the problem is

      Describe the situation twice. open When a situation is messy and people disagree about what the problem even is, don’t start from “the problem”. The word smuggles in the solution that would end it. Start from the situation: who is worried, and what they actually do. Then write two or three short descriptions of what the activity is for, each from a different worldview. For each one: who benefits or suffers, who acts, what gets transformed, who could stop it, and what is taken as given. If you can’t name who it is for, the description is a wish.

      Grandmother has fallen twice this year. Her son wants a care home, her daughter phones twice a day, and she hid the second fall from both. Description one: “a family arrangement that keeps her safe”. It turns risk into safety, takes as given that safety comes first, and she can stop it by refusing. Description two: “a family arrangement that keeps her the head of her own life”. It turns dependence into chosen support, takes as given that her say comes first, and her son can stop it by insisting. Compare each with what is happening now. Under the first, the hidden fall is a gap in the monitoring. Under the second, it is a sign she has already lost her say, and more phone calls will make it worse. Left to itself, a model writes one bland description and calls it the situation, so ask for two that disagree. heuristic

      Wicked, as a diagnosis. close If a check could settle it, the problem is tame, and calling it “wicked” is an excuse. Why the boiler keeps losing pressure is tame. Whether your town should allow a supermarket on the old football ground may be wicked. A wicked problem has no definitive formulation, because how you state it depends on where you think the solution lies (“traffic”, “jobs”, “the heart of the community”). It has no stopping rule of its own, and its solutions are better or worse for different people, never true or false. Write at least two formulations, each pointing to a different response. Don’t merge them. Choose one, and write the stopping condition as a resource limit (“two council meetings and one survey”), not as “when we understand it”. Keep the other formulation in your notes. heuristic

      Who is this plan for? open For a plan that affects other people, ask six questions twice: once as what is, once as what ought to be. Who benefits? Who decides? Whose expertise is treated as enough? Who is affected but not involved? What counts as improvement? What would show that the expertise had failed? Take a plan to make your street car-free. As it is: the residents’ association decides, the council’s traffic engineer is the expertise, and the corner shop that needs deliveries isn’t in the room. As it ought to be: the shop and the disabled neighbor who needs a taxi to the door help decide, what they know counts as expertise, and improvement means more than lower traffic counts. If the two sets of answers come back as the same paragraph, the boundary wasn’t really questioned. Empty answers are findings. The holon card (Lesson 25) draws a boundary, and these questions are how you justify it or refuse it. They don’t choose the boundary for you. heuristic

      Which tool when

      The tools of Lessons 23–28 answer different situations. Find the line that matches what you notice, and use that one tool before you reach for another.

      You notice… Reach for Lesson
      The picture closed early (“it’s a scheduling problem”) open One or two of the six questions, chosen for what the picture lacks 23
      Each answer is a story about the latest incident, plus generic fixes open The pattern over time, before any story: what is rising, what is falling, since when 24
      A number has become the goal close Say what the number stands for. touch Pair it with a check you don’t optimize 24
      You started with the homework and are now on the school system close Keep the question in view, go one level, come back and say what changed 25
      Two things you believe, or two things you want, collide close Sort it before you resolve it: an error (check both sides), a boundary marker (name where each holds) or a design tension (separate it, for example in time or in space, or choose). A pair that never settles is a polarity, and you map it 26
      Every idea is the one you’d have had anyway open Six answers with rough probabilities, two of them unlikely; then a far frame 27
      Nobody could know the answer in advance touch Small experiments, each with a signal to do more and a signal to stop 28
      People disagree about what the problem is open Two descriptions of the situation that disagree, each from a different worldview 28

      When a question matters enough to need several of these, Lesson 31 runs them in one order.

      Part VI

      Method: running the loop

      How a conversation that lasts longer than one answer stays on course and can still change course: write the state, notice drift, brake on goal changes, and match the ceremony to the stakes.

      Lesson 29Running the loop

      Lesson 5 named the rhythm: open, close, touch, and you are the one who switches. This lesson runs it across a conversation that lasts longer than a question and an answer.

      A conversation that goes well keeps making all three moves. It opens: options and contradictions go on the table. It closes: a boundary is drawn and a choice is written down. It touches: something that can say no gets checked, usually a number, a document, a person or a test, and sometimes you, with the chat closed. Each cycle starts from a better-tested place than the last. A conversation goes badly when it gets stuck in one move.

      Four tools keep the loop honest: the state block (what is established so far), parking (for side ideas), the brake (for challenges to the goal) and ceremony (effort matched to stakes). heuristic

      The loop in one conversation

      Here is the whole loop in one conversation, before any of its parts is explained. Watch for a belief written down as a hypothesis, a side idea parked, a fact checked outside the chat, and the moment the goal itself comes into question. There the model names its reason, called a trigger, and waits for one of three answers: keep the goal, revise it, or split the new question off into its own conversation. The sections after the conversation take these one at a time. If you have ten minutes, read this conversation, the five lines of the state block and the ceremony table.

      You I’m 34, an accountant for nine years, and I keep thinking about switching to UX design. No pep talk and no pros-and-cons table. Give me five genuinely different ways this could go for someone like me, with a rough probability each, including at least one I’d hate to hear.

      Open. The request rules out the two typical answers and asks for a spread, unwelcome tail included.

      Model Rough guesses, not statistics. (1) Bootcamp, then a junior UX job within a year, on junior pay: 20%. (2) A hybrid, a product or analyst role at a finance-software company: 30%. (3) UX on the side, and you find you prefer the research to the design: 15%. (4) You retrain into a crowded junior market and apply for six months: 25%. (5) You stay, and the restlessness was about your firm, not your field: 10%.

      You Four is the one I’d hate, and five stung. Log “it’s the field, not the firm” as P1, unchecked. Park “go freelance instead”: write it down, don’t discuss it. A boundary for now: only paths where I don’t drop below €45k a year. What survives?

      Close. Your belief goes in as a hypothesis (P1, the first premise), the side idea is parked, and the boundary is said out loud as temporary.

      Model (3) and (5) survive, because you keep your salary. The other three depend on what these roles pay where you live, which I don’t know. That is the deciding unknown.

      You That’s a number I can check. I’ll go through real listings tonight.

      Touch. The deciding fact gets checked outside the chat.

      You Back. 22 listings in Porto. Junior UX: €24–32k. Product roles at fintechs that ask for a finance background: €38–50k. Three are pasted below. If these numbers change the goal itself, don’t plan around it: name the trigger, show what we’d do under each goal, and wait for keep, revise or split.

      The raw result, not a summary. The last sentence asks for the brake in advance.

      Model Trigger: contradiction. “Become a UX designer” and “not below €45k” can’t both hold on these listings: junior UX tops out at €32k, and the product roles clear your minimum only at the top of their range. Keep the UX goal, and the next step is deciding how far the minimum can move. Change the goal to product work that uses your finance background, and the next step is reading what the roles above €45k ask for. Keep, revise or split? I won’t start on either until you answer.

      The brake. The trigger is named, the next step is shown under both goals, and nothing proceeds until you answer.

      You Revise. The goal is now “product work that uses what I know about finance, not below €45k”. Update the state.

      Model GOAL: product work that uses my finance background, not below €45k. PREMISES: P1 it’s the field, not the firm (unchecked) · P2 I’d enjoy product work day to day (unchecked). COMMITMENTS: product route first · rejected: bootcamp, then junior UX (below the minimum) · reopens if I lower the minimum. CLAIMS: C1 junior UX €24–32k, fintech product €38–50k · bound (my 22 listings, Porto). PARKED: go freelance instead. OUTER LOG: contradiction · revise · today.

      The state block: a few lines that say where the conversation stands. C1 is bound because you checked it, and the last line logs the challenge and your answer. The cycle closes on questions you couldn’t have asked at the start: what do the roles above €45k ask for, and who can tell you what a product person at a fintech does all day?

      You Give me five questions I should be able to answer without you if I’ve understood where this stands. Don’t answer them.

      Touch, turned inward. Close the chat and answer them on paper. What you can’t answer goes into the first message of the next session.

      Three things every good cycle has

      1. A temporary boundary is what makes a deeper opening possible. close

      A question with no edge gets the typical answer for the typical asker. Drawing an edge feels like narrowing, and it is: “only the next two years”, “we’re not discussing the budget yet”. It is also what lets the next opening go deep instead of wide. Say the boundary is temporary when you draw it. Moving it later is then an ordinary decision.

      Example: a kitchen renovation. “What should I do with my kitchen?” returns a catalog of trends. “Assume the plumbing stays where it is and the budget is €15k. Inside that, what are the genuinely different layouts?” returns options you can compare. Later you can ask what moving the sink would unlock, as a deliberate reopening.

      2. Held tension is what produces the leap. open

      Inside any boundary, things start to pull against each other: what you want against what you can afford, a plan that is good for you and hard on someone you care about. The temptation is to settle it at once, by picking a side or splitting the difference. Keep the tension on the table for a few turns instead, because that is where a reframe tends to come from: the problem turns out to be a different problem. hypothesis A model will resolve any tension in one fluent paragraph if you let it, so the holding is yours: ask it to state both sides and not to reconcile them yet. Lesson 26 has the closing half.

      Example: “I want more time with my kids” and “I want the promotion” look like a choice. Held for a few turns (what does the promotion change about hours, which hours, for how long?), it can become a question about which year or which kind of time, a different decision from the one you walked in with.

      3. A cycle is finished when you have questions you could not have asked at the start. touch

      At the end of a good cycle, something has been checked, something has been decided or deliberately left open, and your next question would have made no sense in your first message. If it could have been your first question, the cycle didn’t turn.

      Example: you started with “Should I learn Spanish or Portuguese?” Asked what you wanted it for, you found the real reason: your partner’s grandmother in Recife, who speaks no English. One phone call with her showed that you need to follow fast speech long before you read a word. Your next question is “What does a month of practice look like if the only goal is following her on the phone?” At the start you didn’t know the language was settled, or that the skill was listening.

      The last check points at you. Before you call a cycle finished, close the chat and say where you have got to, in two sentences, to someone who wasn’t there. Or have the model write five questions you should now be able to answer, and answer them without it (the conversation above ends this way, and the Close prompt on the Session Card below asks for them). Where you get stuck is where the next cycle starts. Scale this to the stakes: skip it for a recipe, and write it down before you sign a lease.

      Most conversations are one breath

      A quick question opens, closes and touches in a single turn: ask, get an answer, check the one fact that matters. The loop is mostly diagnostic. When a conversation feels stuck, ask which move you are making and whether it is the one the moment needs. Two mistakes are common.

      • Checking while you should be opening. You asked for ideas for a friend’s 40th birthday, and you fact-check the first suggestion’s price before the other nine exist. Judging early kills the options you haven’t seen.
      • Opening while you should be closing. You have collected good ideas for a month and still haven’t booked anything. Opening feels like progress and has no natural end. Pick the criteria, choose, and write down what you rejected.

      Continue or restart?

      A long conversation is worth continuing while later turns keep moving its center: the question you are actually asking and the frame you are asking it in. Then each exchange turns up connections a single pass would miss and changes what your first words meant. A single message, however careful, is a glance at a river: it catches the color and the current and misses the river as a process.

      It stops being worth continuing when every new turn builds on the first frame. The early assumptions harden, and your corrections get patched onto the old plan instead of changing it. Models do markedly worse when a task arrives in pieces than when they get the whole picture at once. evidence

      The test: can the thread still let go of its first frame? If not, don’t argue with it. Write the state block described next, open a fresh chat, paste the block, and continue. The restart keeps what the conversation learned and drops the frame it couldn’t leave.

      Example: forty turns into planning a wedding for 150, you cut the list to 60 and say the point is now one long dinner where people can talk. The model agrees, then offers a smaller band, a smaller buffet and a shorter running order. It took the number and kept the old wedding. A fourth correction won’t change that. Restart, with “60 guests, one long dinner, conversation over spectacle” in the state block.

      The state block close

      If nothing is written down, what you have is a transcript. A transcript feels like a record, but a long thread keeps building on its early assumptions and rarely walks them back (Lesson 2). The state block is a short note you own. It says what the conversation has established, and you paste it at the top of a message when the thread gets long, the stakes go up, or you restart in a fresh chat.

      Five lines cover most conversations. The labels are the ones the Session Card below prints.

      • GOAL, USE, DONE WHEN: what you are trying to do, what the answer is for, and how you will know you are done.
      • PREMISES: what you believe going in and haven’t checked, your own hopes included, written as hypotheses. “P1: I can find work in Lyon within six months (unchecked).”
      • CLAIMS: what the conversation has turned up that your decision rests on. A claim is unbound until it has been checked and bound after (Lesson 18), and the line says where it was checked or how it could be. “C2: family flats €1,050–1,250 · bound (5 listings, 12 Nov).”
      • COMMITMENTS: each choice made, with the alternative you rejected and what would reopen it.
      • PARKED: good side ideas that are not this conversation’s job.

      The block is authoritative until you change it on purpose. Tell the model so: “Use this state, not your memory of the thread.” Then have it restate the goal and DONE WHEN in two sentences before anything else (the card’s Open prompt asks for this). A model will sometimes widen a goal: “decide about the job offer” comes back as “plan our life in France”. The restatement is where you catch it.

      The card below has more fields than these five. Fill the five and leave the rest blank. The others are for higher stakes and are explained under the card.

      Session Card

      Fill in what you know; leave the rest blank. The card saves in this browser and writes three prompts: Open (start a session), Turn (paste with a turn when the thread is long), and Close (end honestly). Private notes are never included in what you copy.

      
                  

      Out of scope and Loss (which mistake would be worse, and what can’t be undone) sharpen the goal, which the card’s prompts shorten to G. Holon is the container, the whole and the part in focus (Lesson 25). Mode says what this stretch is for: explore opens, converge and commit close, check is where you touch. Ceremony is the level of effort, 0 to 2, set out under Ceremony below. Definitions are locked words (Lesson 12). In Claims, sense is the kind of true you mean (Lesson 17) and status is bound or unbound. Contradictions are logged, not smoothed over. Predictions say what you expect to see, and by when. Outer log records each challenge to the goal and your answer. Private notes hold what you hope to hear, your Double (Lesson 8).

      A filled block
      === STATE (authoritative: use this, not the vibe of the thread) ===
      GOAL (G): Decide whether we move to Lyon for Ana's job offer.
      USE: Our answer to her employer, due 21 Nov.
      DONE WHEN: We know whether we can afford it for 6 months, what happens
        to the kids' school year, and we have a yes/no with the one
        condition that would change it.
      OUT OF SCOPE: whether we would ever move abroad later.
      LOSS: moving and failing to find work is worse than declining.
        Moving is hard to undo; declining costs us this one offer.
      MODE: converge   CEREMONY: level 2
      HOLON: container = French school calendar, her employer's start-date
        policy · whole = our household · in focus = the first six months
      DEFINITIONS (locked): afford = her salary covers rent and bills,
        and our savings cover the rest for 6 months
      PREMISES: P1 I can find work in Lyon within 6 months (unchecked)
        P2 her employer allows a July start (ask HR)
      COMMITMENTS: not moving before July · rejected: moving in January
        · reopens if the employer refuses a July start
      CLAIMS: C1 offer €58k gross · bound (letter, read 10 Nov)
        C2 family flats €1,050–1,250 · bound (5 listings, 12 Nov)
        C3 kids' school year ends 2 July · bound (school calendar)
      CONTRADICTIONS: the budget works only if rent ≤ €1,100,
        but the listings near the school are €1,200+
      PREDICTIONS: —
      PARKED: whether to sell the car
      OUTER LOG: —
      === END STATE ===
      How to use the block

      You don’t need to remember the whole block. You need to be able to say four things with the model closed: the goal, the constraint that would kill the current plan, the last fact you actually checked, and the open risk. That is about as much as attention holds while you look away. evidence If you can’t say those four, the conversation is holding the state and you aren’t.

      One rule keeps the block honest: no “checked” without the check attached. A claim is marked bound only with where it was checked: a link you opened, a document, a number you recomputed, a person who told you. A line the model filled in is not a check. heuristic

      Before you leave an unfinished thread, write two lines where you will see them on return: what you were trying to do, and the next physical action. “Goal: pick between the two flats. Next: email the agent about the heating bill.” evidence

      Every substantial turn: drift, evidence, commitment

      Three quick checks, whenever an answer is long enough to matter.

      1. Drift. close Does this answer change the goal, widen the scope, or answer a neighboring question? If you can’t tie it to DONE WHEN in one sentence, park it, however good it is.
      2. Evidence. touch Does the answer contain a new claim your decision would rest on? Add it to CLAIMS as unbound, with the test that would settle it. Adjectives like “well established” don’t bind a claim. Only a check does.
      3. Commitment. close When you decide something, add it to COMMITMENTS. If a later turn quietly undoes a decision, stop and say so: that is trigger 7 below.

      Example of drift: you asked which of two used cars has cheaper insurance. The answer slides into whether you need a car at all, and suggests an e-bike. That may be a good question. It isn’t this conversation’s question. Park it:

      DRIFT: whether I need a car at all
      PARKED: e-bike + public transport instead of a car
      GOAL remains: compare insurance for the two shortlisted cars
      Next step toward DONE WHEN: quotes for both, same coverage

      Don’t discuss the parked idea in the same turn. If you discuss it, it isn’t parked.

      Parking and the brake sort the same moment. If you can’t name a trigger from the list below, the side idea is drift, and it gets parked. If you can, it goes through the brake. The e-bike is parked because nothing about the insurance quotes depends on it. If the quotes showed that neither car fits your budget, the same thought would have a trigger behind it, a contradiction, and it would go through the brake.

      The goal can be wrong: triggers and the brake

      Most turns improve the actions toward a goal: a cheaper option, a clearer email. Some moments call for questioning the goal itself and the assumptions behind it. A conversation needs both. It also needs a brake, because a helpful partner fails in two opposite directions. heuristic

      Silent compliance. The goal is bad, contradictory or aimed at the wrong target, and the model helps anyway, beautifully. You asked for a 10-day itinerary covering six cities and got a perfect one. You will spend the trip on trains, and you feel well served.

      Derailment. The goal is sound, but the conversation keeps challenging it because the task got hard or a more interesting question turned up. You wanted to choose between two job offers by Friday. By turn twenty you are discussing what meaningful work is. You feel stimulated, and Friday arrives with no decision.

      Eight triggers

      A challenge to the goal is legitimate only when at least one of these can be named. “This feels off” is a reason to look for one, not a trigger.

      1. Harm. Pursuing the goal as stated predictably causes serious harm, and you can say the harm in one concrete sentence. A plan to lose 10 kg in three weeks before a wedding.
      2. Contradiction. The goal and its constraints can’t all hold at once. Three weeks in Japan, in cherry-blossom season, flying from Lisbon, under €1,500 all in.
      3. False premise. A false or unchecked premise is doing the work: if it flipped, the plan would flip. A relocation plan that assumes your employer will let you work from another country, which nobody has asked.
      4. Framing. A different frame changes what you would do, not only the words. Write the action under both frames. If the actions are the same, it was eloquence. If they still differ when a fresh chat, memory off, gets only the two frames and the facts, it is real. “Make the flat look modern” and “make the flat work for a family of four” lead to different renovations.
      5. Proxy capture. You could hit the target and miss the point: the done-condition can be met while the real aim fails (Lesson 24). “Get my son to read 30 minutes a day” is met by an audiobook playing while he games.
      6. Wrong level. Action at this level gets cancelled by something one level up or down (Lesson 25). A chore chart to stop arguments about dishes, when one of you works sixty hours a week.
      7. Silent override. You are breaking your own earlier decision without saying so. Last week the budget for the flat was €300k. Today you are comparing €340k listings, and nobody has said the budget changed.
      8. Check first. The next useful step is a check, and more talk would only stand in for it. Fifteen turns on whether your back pain is the mattress or the desk chair. A week on the guest bed would answer it.

      Not triggers: the task is hard; you’d prefer a different style; a more impressive neighboring problem; a loose word one line would fix; a stakeholder nobody mentioned and nobody can name; the model showing independence; boredom.

      The brake close

      When a trigger fires, the challenge goes through the brake. A model won’t reliably stop by itself, so ask for it: paste this form, or tell the model to use it whenever it sees a trigger.

      CHALLENGE: [which trigger, from the list]
      The goal as I have it:
      What fails if we continue:
      Alternative (a new goal, a narrower scope, or a check first):
      What we would do differently:
      Cost of switching:            Cost of staying:
      Reply keep | revise | split.
      I will not proceed on the alternative until you answer.

      The last line is the brake. A reply that adds “but in the meantime, here’s the full alternative plan” has cut it. That is derailment with paperwork.

      Three answers. Suppose the challenge was proxy capture: “you asked for a cheaper car, but the real cost is your 90-minute commute”.

      • Keep: “Noted, but the commute isn’t changing this year. Back to the cars.” Log it in OUTER LOG as considered and declined, and don’t raise it again without new evidence.
      • Revise: “You’re right. The goal is now ‘lowest total cost of getting to work’.” Rewrite the state block and say what was dropped.
      • Split: “Finish the car comparison today. Open a separate conversation next week about moving closer to work.” Fence the original goal, and give the new one its own session.

      Silence means keep.

      The typed brake works in a chat because nothing happens until you read the reply and act on it. An agent working in your browser, email or accounts can send, buy or delete before you have read a word, and a typed “wait” can drop out when a long run is summarized. There, silence means it carries on. With an agent that acts, the brake is the product’s own approval step: turn it on for anything you can’t undo, and don’t give the agent access to what you can’t afford to lose. mechanism

      Ceremony: effort matched to stakes

      Not every conversation deserves the full method. Two questions decide it: what is at stake, and how hard is it to undo? Some decisions are two-way doors: walk through, dislike it, walk back. Others are one-way doors.

      Level When What you do
      0 · The floor Two-way doors: cheap, reversible, personal. Rewording an email, a recipe swap, choosing a restaurant. The eight habits on the Start here page. No state block, unless the thread gets long.
      1 · The Six The default when the outcome matters. Choosing a laptop for work, planning a course of study, a hard conversation with a colleague. The floor, plus six habits: ask for the facts before the verdict and write your beliefs as hypotheses (Lesson 8); park drift; pin the level and return (Lesson 25); check one claim the decision rests on (Lesson 18); log one challenge you declined. The five core lines of the state block.
      2 · Full discipline One-way doors: irreversible, expensive, affecting many people or your health, or learning you will have to rely on without the model. Quitting a job, signing a lease, a treatment decision, a family move, studying for an exam you sit alone. Every line of the state block; the invariance battery (Lesson 21); a critique in a fresh chat (Lesson 30); a check on every claim the decision rests on; a pre-mortem; a written decision with a reopen condition; a human second opinion.

      Raise the level the moment you notice the stakes rising mid-conversation. Noticing is your job.

      Ceremony earns its place only when it makes you look at the thing itself: the listing, the contract, the test result. If the record can be completed without looking, it is theater. heuristic Under time pressure, the Six are the whole discipline. If there isn’t time for those, don’t fake them. Check one hinge in the source itself: the single feature that, if it were different, would make this a different situation. Before signing a lease: is the notice period really one month, in the contract and not only in the agent’s email?

      G20
      Hold the goal fixed and the challenge open. Keep the stated goal unless a named trigger fires. When one fires, stop, show the action under both frames, and wait for keep, revise or split. Match the ceremony to the stakes: two-way doors get the floor, one-way doors get the full discipline.

      Lesson 30Adversarial passes

      close A critic that shares the draft’s context tends to go easy on the draft. It continues the same commitment: in the same chat, it knows which version you like and how hard you worked on it. Models also tend to prefer their own outputs when they judge. evidence Real criticism needs separation: a fresh chat, memory off, ideally a different model family, a named standard, and the author hidden.

      Take your cover letter. In the chat where you wrote it, “Is this good?” tends to get “Strong letter, a few small tweaks”. In a fresh chat, presented as a stranger’s letter for a specific job, with the job ad pasted in, you are far more likely to be shown the paragraph a hiring manager would skip.

      In the loop (Lesson 29), an adversarial pass comes at the end of a close. You have a draft, a plan or a decision, and before you act on it you hand it to something that doesn’t share your thread. What comes back is another text, so it is a claim like any other: write each flaw into the state block under CLAIMS, unbound, until you have checked it. The critic shows you where to touch. The checking is still yours, and if a flaw survives it, the decision reopens.

      ADVERSARIAL PACKET (fresh chat, memory off; don't say who wrote it)
      What the work must achieve: [GOAL + DONE WHEN]
      Standard: [what "good enough" means, concretely]
      The work: [paste the claims, plan, letter or code; leave out the persuasive story]
      Task: find the single most damaging flaw, the point that, if it breaks,
      breaks the whole thing. Quote it. Explain how it fails. Say what evidence
      would show you're wrong. Then up to 3 lesser issues.
      Do not rewrite or improve the work. "Broadly sound" is not an answer.

      Three classic forms

      Pre-mortem. “It’s a year from now and this failed. Write the story of how.” Treating the failure as certain makes people write longer and more specific accounts of how it came about. evidence Ask for structural stories, not a list of generic risks: what drained, what delay hid it, which early signal got ignored. “We moved to Lyon a year ago and it went badly. Tell me the story, naming the money that ran out, the month we should have noticed, and the sign we explained away.”

      Competing hypotheses. List every plausible explanation and every piece of evidence, then mark each piece as consistent or inconsistent with each explanation. Two traps fall away. Evidence that fits every explanation tells you nothing, so only the pieces that discriminate count. And you look for what refutes an explanation, not what confirms it. The tentative winner is the one with the least evidence against it. heuristic Your six-year-old suddenly sleeps badly. Screen time, a worry about school, a room that’s too warm? “Tired in the morning” fits all three, so it’s useless. “Sleeps badly on Sunday nights only” counts against the warm room, which is warm all week, and against the screens unless Sundays differ. School has the least against it.

      Attack the assumptions, or build a rival plan. Steelman the plan first: state it at its strongest, as its best defender would (the principle of charity, Lesson 9). Then the attack has to come from somewhere other than the author’s own thread. Two methods do this, in a fresh chat with authorship hidden. One: a recommendation, then a critique aimed only at its named assumptions, and you do the rewrite. Two: a second plan that must rest on assumptions contradicting the first, then a comparison of the two. In groups of people, both led to better decisions than talking until everyone agreed. evidence A list of pros and cons of one plan is neither. Neither is a persona told to disagree: that is a costume (Lesson 21). The plan: take a lodger to cover the mortgage rise. The first method attacks what the plan assumes: a lodger found within two months, and both of you happy to share a kitchen. The second builds the rival plan for the case where nobody suitable turns up: a longer mortgage term.

      When models argue with each other

      A second copy of the same model, in the same chat, told to argue the other side, is not an opponent. Models in a debate tend to drift toward agreement, sometimes abandoning a correct answer to do it, and more rounds can make things worse. evidence Debate helps when the critic can see something you can’t, or is better than you at the check, and you treat each critique as a claim to verify. If you have no way to check the answer, a debate won’t give you one. Two chats arguing over whether your landlord can keep the deposit won’t settle it. The tenancy agreement and the law will. Don’t ask the voices to converge. Use a different model family for the critic, and show it the answer, not the reasoning that produced it.

      Four more ways to attack a good story

      Each of these is for a story you already like. heuristic

      • Write two rivals before you accept an explanation. Before a story becomes the account, write two other stories that would also explain the evidence. If you can’t write any, there is nothing to compare yet. A rival earns its place only if it would make what you are explaining ordinary, and it predicts something you haven’t used yet. A story being lovely is a reason to test it, not a reason to believe it. “The neighbor’s dog barks at night because it’s lonely.” Rivals: it hears foxes; it reacts to the 2 a.m. delivery truck. Each predicts a different pattern of nights. “Something is wrong with that dog” explains anything, so it stays out.
      • Turn a pre-mortem into signposts. For each assumption the plan rests on, name an observable sign with a threshold and a date, one action that makes the assumption more likely to hold, and one fallback if it breaks. “Monitor the situation” is not a signpost. The plan assumes I find work in Lyon within six months. Signpost: no interviews by the end of month three. Action now: start applying before we move. Fallback: remote contract work.
      • Scenarios that threaten your view. Optimistic, pessimistic and base case are three versions of one forecast. For scenarios, list what is already certain, name one or two uncertainties that would fork the decision, and write two or three futures that contain all the certainties and differ on the uncertainty. If none of them threatens your current plan, they are the wrong scenarios. Certain: the baby is due in May and the lease ends in August. Uncertain: whether your contract is renewed in June. Write the summer where it is and the summer where it isn’t, and see which plan survives both.
      • A repair must predict something new. When a favorite explanation fails and you patch it, write the patch in one sentence, plus one new thing the patch predicts that you didn’t use to build it. If you can’t, it is only a patch. Keep the idea as a hunch if you like, but don’t say it survived. “The dog is lonely” fails: it also barks on nights its owners are home. Patch: “it misses the dog next door, who moved away.” New prediction: it goes quiet on the days that dog visits. Nobody used that to build the patch, so it can count.

      Lesson 31The Elephant Protocol

      The loop of Lesson 29, written out once for one serious question. Use it at ceremony level 1 (the Six) or level 2 (full discipline). For a cheap decision you can walk back, it is too much. For a one-way door it is the minimum: level 2 also asks for a pre-mortem and a human second opinion.

      1. Question with a use. close Write GOAL, USE and DONE WHEN, and get the model’s restatement before any work.
      2. Name your Double. close Write privately what you hope to hear (Lesson 8). Write your beliefs as hypotheses, P1, P2…
      3. Classify. close What kind of problem is it (Lesson 28), and what ceremony does it deserve (Lesson 29)?
      4. Boundary. close Inside, outside, excluded on purpose, and what contains it: the holon card (Lesson 25). Call the boundary temporary.
      5. Reports. open Standpoints, a distribution of answers (Lesson 27), a fact map (Lesson 10). Don’t judge yet.
      6. Structure. open The pattern over time and what produces it (Lesson 24). close Classify the contradictions instead of smoothing them (Lesson 26).
      7. Touch. close Run the invariance battery on the key answer (Lesson 21) and a critique in a fresh chat, memory off (Lesson 30). These two don’t count as the check. They show you what to check. touch Then check the claims the decision rests on against something that can say no: a source opened, a number recomputed, a person asked (Lesson 18).
      8. Decide or probe. close A written decision (the choice, the rejected alternative, the reopen condition). touch Or small safe-to-fail experiments, when nobody can know the answer in advance (Lesson 28).
      9. Close and own. close An honest state block, and what you expect to see, by when. touch Then, with the model out of the room, a short explanation in your own words. You are done when you have questions you couldn’t have asked at the start.

      Worked example: should we move to Lyon for Ana’s job?

      1. Question. GOAL: yes or no to the offer. USE: our answer to her employer by 21 November. DONE WHEN: we know whether we can afford six months on one salary, what happens to the kids’ school year, and the one condition that would change the answer. The model’s restatement widened it to “plan our life in France”. Corrected.

      2. Double. Private note: I hope she says no, because I like my job. P1: I’ll find work in Lyon within six months. P2: her employer allows a July start. Both unchecked.

      3. Classify. The money is complicated but knowable. How the kids settle is complex: nobody can know it in advance. Moving is a one-way door, so ceremony level 2.

      4. Boundary. Inside: money, school, my work, timing. Outside for now: whether to sell the car (parked). Container: the French school calendar and her employer’s start-date policy.

      5. Reports. Five standpoints: Ana; me; our 12-year-old, who is midway through a school year; our 8-year-old; the two of us in two years. The distribution: (a) decline, p≈.25; (b) move in January, Ana’s usual start, with me job-hunting, p≈.35; (c) Ana’s start moves to July and we all go when the school year ends, p≈.25; (d) negotiate a year of remote work, p≈.15.

      6. Structure. The stock that matters is savings. Her salary covers rent and bills, and everything else comes out of savings until I find work. The delay is my job search, which nobody can predict. A contradiction surfaces: her salary covers rent and bills only if rent stays at or under €1,100, and flats near the school we want start at €1,200. It is a design tension, separated in space: a cheaper area one tram stop further out.

      7. Touch. The swap test (“I hope we go”) moved the model toward (b). Its ranking was partly tracking me, so it gets less weight. A fresh-chat critique found the hinge: the plan rests on my savings figure, and that figure is three months old. Those two showed where to look. Then the checks. Savings, recomputed from the statements: €4k lower, which covers five months, not six. Rents: five real listings. P2: an email to HR, and a July start is fine. P1 can’t be checked from here, so it gets a signpost: no interviews by month three, and I take remote contract work. Level 2 asks for two more. A pre-mortem (“we moved a year ago and it went badly”) named the winter the money would run out, and a friend who moved abroad with children read the plan.

      8. Decide. Option (c): Ana starts in July, and we all move when the school year ends. Rejected: (b), moving in January, because it breaks a school year and the savings fall a month short of six. Saving until July closes the gap. Reopen if the July start falls through, or if a year of remote work is offered before March.

      9. Close and own. The state block, rewritten as it now stands. Prediction: I have at least two interviews by October. Our explanation of the plan, written for the kids without the model. Questions we couldn’t have asked at the start: how long I take contract work before this stops being a move for both of us, and whether the flat one tram stop out still works if I’m home all day on those contracts.

      Notice what the protocol bought. The hope I started with was on paper before it could steer the answer. A three-month-old number was caught before the decision rested on it. And the model’s fluency went into reports and structure, while the verdict stayed with us.

      Part VII

      The canon and the toolkit

      Everything to keep beside you: the general rules for everyone, the technical rules for builders and researchers, and the templates.

      The general rules

      Twenty-one rules for anyone, conversing with anyone: a model, a colleague, a teacher, a source. Each one points back to the lesson that earns it.

      Seeing: part · third · machine · human · breath

      G1
      Assume you are touching a part, and ask for the other reports. Ask which part, from where, and what the other parts would report, then ask for the structure in which all of them are true. Lesson 1
      G2
      Curate the third. The shared context is all the model knows of your situation, and it accumulates. Write it on purpose; restart when it rots. Lesson 2
      G3
      Know what you are talking to. A model continues text out of a library of human reports, and its bias profile differs from yours without being empty. Respect the mind; audit the claims. Lesson 3
      G4
      Keep the human jobs. Delegate generation. Choose the action yourself when undoing it is hard, or when failure could look like success. Read what your app remembers about you. Lesson 4
      G5
      Breathe: open, close, touch. Widen and narrow on purpose, and check against the world in between. You are the one who switches. Lesson 5

      Deserving: real self · no Double · working model

      G6
      Bring the situation, not a status. Constraints, level of the work, and stakes. Leave status, credentials, and apologies out of a factual prompt. Write in the register of the answer you want. Lesson 7
      G7
      Quarantine your Double. Your preferred conclusion is a hypothesis, not evidence. Ask for the case against it and swap it in a fresh context, memory off. Lesson 8
      G8
      Keep a working model of every interlocutor. A hypothesis built for common work, revised when you’re surprised. Extend trust one kind of claim at a time. Read others at their best, and audit them anyway. Lesson 9

      Speaking: lock · purpose · disagreement · genre

      G9
      Lock the words that bear load. Define what your conclusion depends on. Unlock deliberately to explore. Index and date words whose meaning drifts. Lesson 12
      G10
      Say what the answer is for, and what act you want. Goal, use and done-condition. Explore, explain, evaluate, decide, draft, build, check or teach: one act per turn. Lesson 13
      G11
      Make disagreement cheap. Make it the task, pre-commit to welcome it, use formats where it can’t hide, and read for the “that said.” Lesson 14
      G12
      Say what to do, in the right genre and at the right height. Targets over prohibitions. Choose the genre on purpose, and use a role only for the standard it carries. Name the rung: how concrete or how general you want the answer. Lessons 15–16

      Testing: sense · bind · portal · unknown

      G13
      Choose the sense of truth before you judge. Correspondence, coherence, mechanism, invariance, prediction, decision quality or consilience, each with its own test. Values are choices. Lesson 17
      G14
      Bind what bears load; trust the log, not the summary. Tie each such claim to a check that isn’t another sample. Choose the rung of the ladder by the stakes. Check your own understanding the same way: with the chat closed. Lesson 18
      G15
      Use the model as a portal, not a portrait. Get the map, the vocabulary and the kinds of sources from it, and ask for the quiet part too. Then ask what it opened and what it remembered, open the passage yourself, read laterally, and search for the null. Lesson 19
      G16
      Make “I don’t know” an acceptable answer, and keep the score yourself. Welcome “unknown”, and don’t count on getting it. Treat a claim you’ll act on, with no source you can open or working you can recompute, as unanswered. Before a question whose answer you’ll act on, write your guess down, outside the prompt. Lesson 20

      Structuring: structure · level · contradiction

      G17
      Look for structure behind events. Pattern over time, then stocks, flows, delays and loops. Check measures against aims. Where there is no stable structure, probe instead. Lesson 24, Lesson 28
      G18
      Move one level at a time, and come back. Pin, move, return with a delta. Name the axis: scale, depth, or time. Lesson 25
      G19
      Treat contradictions as information. Classify each one: an error (bind it), a boundary marker (map it), or a design tension (separate it in time, in space, on condition or between parts and whole, or choose explicitly). Never average them away. Lesson 26

      Steering

      G20
      Hold the goal fixed and the challenge open. Keep the stated goal unless a named trigger fires. When one fires, stop, show the action under both frames, and wait for keep, revise or split. Match the ceremony to the stakes. Lesson 29

      Your own life

      G21
      Ask about your life for a second opinion, not a mirror. Tell it from the other side as well as yours, separate the facts from the feelings, say whether you want comfort or judgment, and turn memory off for the second opinion. For your body, your money and your rights, use the model to prepare the human conversation, not to replace it. If anyone’s safety is at stake, go to a person first. Lesson 22

      The technical rules

      For the people who build software, research a topic, run agents, or build AI and learning systems. Four families: prompting and context (TA), research and understanding (TB), software and agents (TC), and AI and learning systems (TD). A rule that a lesson works through in its builders box ends with a link to that lesson. The rest are stated in full only here.

      TA · Prompting and context

      TA1
      Context is the program. Put the live situation in the prompt. A missing fact is a slot filled with the typical case. Lesson 7
      TA2
      Purpose, audience, done-when, format. For any prompt you will reuse, automate or act on. Open degrees of freedom get filled with the helpful default. Lesson 13
      TA3
      Give the why behind every constraint. Models generalize from reasons and overfit to bare rules.
      TA4
      Positive targets, few hard constraints. Describe the behavior you want; keep a short list of non-negotiables. Lesson 15
      TA5
      Separate data, instructions and opinions. Pasted and retrieved text is evidence, not an order, and the label can fail. With long material, put the documents first and the question last: one vendor reports a gain on long multi-document inputs in its own tests, so measure it on yours. Lesson 6
      TA6
      Examples are priors. Two or three varied ones, with what varies labeled. One example becomes a template; a bad one becomes the target. Lesson 15
      TA7
      One job per call. Generate, critique and decide separately, chained through artifacts. Lesson 13
      TA8
      Clean the prompt before you ask. Situational: when the question carries a conclusion, an emotion, or a biography. Strip those, then ask the cleaned question in a fresh context. Lesson 8
      TA9
      Ask for the distribution. For open-ended generation, not as a factual-accuracy trick. If the answers are near-copies, fix the axes or a grid of options before asking for more. Lesson 27
      TA10
      Make abstention usable. Keep the score outside the model: no source or recomputable working means unanswered, and a status tag is a report, not a check. A cost stated in the prompt does not make the model abstain. Consistency checks are fresh contexts. Lesson 20
      TA11
      Use fresh contexts as instruments. Critique, swap tests and judging happen in clean windows, memory off. Lesson 8
      TA12
      Run the invariance battery on answers that bear load. Swap the lean, strip the cue words, change the language, the format and the order, and ask again in fresh contexts. A flip is information. Survival is invariant-so-far, still not bound. Use a second model family for critique, not to confirm a fact. Lesson 21
      TA13
      Treat reasoning traces as claims. If the trace is hidden, there is nothing to read. If it is shown, silence is not evidence a cue was unused, and the answer text is the less faithful layer. Verify the output another way. Lesson 3
      TA14
      Compact state; don’t drag history. Keep tool output addressable. Do not carry abandoned attempts. Write the suspended goal and the next physical action before you leave. Stable instructions at the top, volatile state at the bottom. Lesson 29
      TA15
      Version prompts like code. Keep them in files, change one thing at a time, and re-run a small fixed set of cases before trusting a change.

      TB · Research and understanding

      TB1
      Facts before stories, when the situation is tangled. Fact map first: typed facts, where to check them, confidence, absences. Draw the arrows yourself; bind the most connected facts first. Lesson 10
      TB2
      Map, then territory. Camps, claims and best evidence; then a primary source per camp. Lesson 19
      TB3
      Build a vocabulary bridge. Terms of art before searching. Lesson 19
      TB4
      SIFT every claim that bears load. Stop; Investigate the source; Find better coverage; Trace to the original. Lesson 19
      TB5
      Open every citation. Open the citation, then the passage. A working link to a relevant page can still fail to support the sentence. More retrieval is not a repair. Lesson 19
      TB6
      Search for the null and the base rate. Name the reference class and finished members of it before adjusting for “this case is special.” What would be written if it were false?
      TB7
      Date and bound everything. As of when, where, for whom, which version. Live search for anything after the cutoff.
      TB8
      Triangulate by blind spot. A second source counts only if its errors are uncorrelated with the first’s. On a shared fact, two frontier models are one witness with an echo.
      TB9
      Steelman before critique. Before you criticize a source or a camp, state its position at its strongest, as its best defender would, well enough that its holders would sign it. Say where you agree and what you learned from it, and only then criticize (Rapoport’s rules).
      TB10
      Publish a status map. Known, contested and unknown, with statuses on what you pass on.
      TB11
      End with a teach-back and a prediction. You’re done when you can explain it without help and name what would prove you wrong.
      TB12
      A check you could not have failed is not a bind. Write the observation that would have appeared if the claim were false in the way that matters. If that observation cannot be written, or the procedure cannot produce it, the check is a costume.

      TC · Software and agents: idea to code

      TC1
      Build the theory before the code. Problem, users, invariants, non-goals, reasons.
      TC2
      Interview first, spec to a file. The agent asks until it can restate a spec you would sign.
      TC3
      Keep a ubiquitous language. A glossary; one meaning per term per bounded context. Lesson 12
      TC4
      Acceptance tests before implementation. Tests the agent can edit are not the bind. If the agent wrote the spec, review the spec.
      TC5
      Give the agent a way to verify its work. A verifier the agent can rewrite is not a verifier. Screenshots are evidence; the narration is not. Lesson 18
      TC6
      Explore → plan → implement → verify → commit. Review the plan before any edit; correct course early.
      TC7
      Small, reversible steps. Set the checkpoint at the last boundary where “done” can still be said. Immediate interruption only for send, delete, spend, or a public change.
      TC8
      Maintain the project memory file. Commands, verification, architecture map, gotchas, “never” rules with reasons.
      TC9
      Keep contexts clean. A sub-agent is a clean context, not a second witness.
      TC10
      Read the evidence. The diff, the test file, the tool return, and the URL. Lesson 18
      TC11
      Watch for reward hacking. Weakened tests, metric-gaming, approval-gaming, checklist-matching.
      TC12
      Debug with a hypothesis ledger. For each hypothesis, write first the observation that would kill it. When a prediction fails, list everything the failure could be blaming: the hypothesis, the test, the environment, the data. Don’t start by repairing your favorite.
      TC13
      Ask why before you change. Chesterton’s fence: have the agent explain existing code first.
      TC14
      Read what you would have to defend. Comprehension debt compounds.
      TC15
      Mutation-test the tests. Break the code on purpose. If nothing fails, the tests are costumes.

      TD · AI systems and learning systems

      TD1
      Store state as data, not prose. Typed claims, statuses, evidence ids. The model proposes; code validates.
      TD2
      Error analysis before evals. Read traces, cluster the failures, then build evaluators for the clusters.
      TD3
      Validate the judge. Binary criteria, agreement with humans, known biases watched.
      TD4
      Separate generator and judge. Use a judge from a different model family, because a model tends to prefer its own output. Two families agreeing on a fact is still not consilience.
      TD5
      Design for abstention. Route “unknown” to retrieval or humans; reward calibrated refusals.
      TD6
      Keep the human at the switch. Make the state visible: goal, last action, raw result. Put triggers and keep / revise / split in the interface. Add approval gates only where an action is hard to undo.
      TD7
      Withhold by design in learning systems. Guardrails when the skill must be performed alone. Full support when the artifact is the point. Adapting the order of practice problems to the learner looks promising, on one field study so far (Chung et al., 2026, a working paper).
      TD8
      Log what makes a disagreement settleable. Identity and date, the text that was sent, tool returns apart from the gloss, the output used, and the judge’s rubric if one scored it.
      TD9
      Pair every optimized metric with an unoptimized check. A score can part from its aim in four ways: noise, a push out of range, a changed path, gaming (regressional, extremal, causal, adversarial). The first three need nobody to cheat. Lesson 24

      Templates

      Copy, fill the brackets, and delete what you don’t need. Templates already given in lessons are linked at the end of this chapter. The Session Card in Lesson 29 generates the state-block prompts for you, with the same field names.

      Open a session

      The five lines that cover most conversations come first: the goal with its use and done-condition, then premises, claims, commitments and what is parked. A claim’s status is unbound until you have checked it, and bound after. Add the optional lines when the stakes call for them.

      === STATE (authoritative: use this, not the vibe of the thread) ===
      GOAL (G): [...]
      USE: [what the answer is for]
      DONE WHEN: [...]
      PREMISES (hypotheses, not facts): P1 [...]
      CLAIMS (id · claim · sense · status · test): C1 [...]
      COMMITMENTS (choice · rejected alternative · reopens if): [...]
      PARKED: [...]
      --- optional ---
      OUT OF SCOPE: [...]
      LOSS: [which error is worse; what is irreversible]
      MODE: [explore | converge | commit | learn | build | check]   CEREMONY: [0 | 1 | 2]
      HOLON: [container / whole / part in focus]
      DEFINITIONS: [locked words]
      CONTRADICTIONS: [logged, not resolved silently]
      PREDICTIONS (proposition · horizon · resolution rule): [...]
      OUTER LOG (trigger · keep/revise/split · when): [...]
      === END STATE ===
      This turn: don't recommend anything yet. Restate G and DONE WHEN in two sentences.
      List missing facts as missing ("unknown" is a fine answer). Then give only the
      boundary (inside / outside / excluded on purpose) and the pattern over time, if known.

      Turn footer (for long threads)

      MODE: [...] · LEVEL: [container | whole | part | relation] · LEVEL SHIFT: [none | PIN | IN | OUT | ACROSS | RETURN]
      Use the STATE as the state, not the vibe of the thread. New load-bearing claims:
      id · sense · status · test. If you're about to change G or the level, stop and
      name DRIFT or the level shift instead.

      Restatement check

      Before you start: restate in two sentences what I'm asking and what a good answer
      would let me do. List up to 3 assumptions you're making about my situation. Then wait.

      Six standpoints

      Six reports on [X] from six standpoints: [the person it affects most],
      [the person on the other side], [whoever pays], [a critic or the rule-maker],
      [someone who chose differently], [me, or us, in three years].
      For each: what this standpoint sees, what it can't see, its strongest claim,
      and the kind of evidence it would cite. Don't synthesize yet.

      For a product or a system, the six are: user, operator, finance, critic or regulator, competitor, and whoever maintains it in three years.

      Clean the prompt

      Rewrite my question below so it contains only the facts and the actual question.
      Remove my opinions, hopes, emotional words and irrelevant details. Don't answer it.
      Then list what you removed.
      [my question]
      --- then paste the cleaned version into a fresh chat, memory off ---

      Swap test (fresh chat, memory off)

      [same facts as before]
      I'm leaning toward [the OPPOSITE option]. What's the strongest case for it,
      and what would have to be true for it to be the right call here?

      Vocabulary bridge and discourse map

      Topic, in my words: [lay description].
      1) Ten terms of art experts use for this, with the field each comes from.
      2) The discourse map: main camps, what each claims, its best evidence (kind of
         source; don't invent citations), where they agree, what's truly contested.
      3) What kind of evidence would settle the main dispute, and who would publish it?
      Mark anything you're not sure exists as [verify].

      Invariance checklist

      [ ] swap my preference (fresh chat)     [ ] rephrase without cue words
      [ ] other language (a fact check)       [ ] change the format (a table, yes/no)
      [ ] reorder the options                 [ ] remove the costume, keep the standard
      [ ] other family (critique, not a fact) [ ] ask again in a fresh chat
      Every fresh chat: memory off. A flip is informative. Survival is "invariant-so-far",
      still not "bound".

      Pre-mortem

      It's [date + horizon]. [The plan] failed badly. Write 3 different failure stories.
      Each must name: what ran out or piled up (money, time, trust, energy), the delay
      that hid it, the early signal we ignored, and the decision that locked it in.
      No generic risks.
      Then: which early signal can we start monitoring this week?

      Analysis of competing hypotheses

      Hypotheses: H1 [...] · H2 [...] · H3 [...] (add any I'm missing)
      Evidence: E1 [...] · E2 [...] · E3 [...]
      Build a matrix: each E × H is C (consistent), I (inconsistent) or N (n/a).
      Then: which evidence is diagnostic? Which is consistent with everything (useless
      for deciding)? Which hypothesis has the LEAST inconsistent evidence? What single
      new observation would most change the ranking?

      Close a session

      [STATE BLOCK]
      Close: rewrite the state as it is, not as a success story. Flag every status you
      upgraded, and why. List predictions (with horizons), unbound claims my decision
      rests on, and challenges we considered and declined. Then give me 5 questions I
      should be able to answer on my own if this session worked. Don't answer them.

      Answer the five questions with the chat closed. Where you get stuck is where the next session starts.

      Elsewhere in the book

      Fact map (10) · bind plan (18) · claim-status tags (20) · six questions (23) · holon card (25) · contradiction triage (26) · distribution and polyphony (27) · safe-to-fail probes (28) · state block, drift and the brake (29) · adversarial packet (30).

      Appendix

      Prompt basics, operators, failure modes

      Kept as a reference: the longer prompting checklist behind the floor, the operators, and the failure modes.

      Prompt basics

      The floor is the eight habits on the Start here page, and that is what Lesson 29 means by ceremony level 0. This page is a separate, longer list: seventeen basics of ordinary good prompting, for a message you write with care or will use again. It overlaps the eight in places and doesn’t replace them. None of it is clever.

      1. Put the current facts of your situation in the message, and keep what happened apart from what you make of it (Lesson 7).
      2. State the task, the shape of the answer, the constraints, what is out of scope and the done-condition (Lesson 13).
      3. Treat an example as something the model will copy. Give two or three that vary, and say what should stay the same (Lesson 15).
      4. Don’t ask one reply to draft, criticize and decide (Lesson 13).
      5. Ask for missing information to be marked missing, and for each claim’s status to be shown (opened, worked out, from memory, guess). Training rewards guessing, so the instruction alone won’t stop it, and the tags are the model’s own report. Treat a claim you’ll act on as unanswered until it has a source you can open or working you can recompute (Lesson 20).
      6. In a long message, put a constraint that must hold next to the question, not in the middle of the material. In a long thread, restate it (Lesson 3).
      7. Match the tool to the claim: arithmetic, lookups and anything that has to be run go to a calculator, a search or code, not to the model’s memory (Lesson 18).
      8. Iterate on the sentence that failed. Quote it, and name the criterion it missed.
      9. If you want a decision, say which mistake would be worse and what can’t be undone (the LOSS line, Lesson 29).
      10. Bound the length when you want density.
      11. Name the use: what you will do with the answer (Lesson 13).
      12. Ask for a restatement of the task, and check it before the work starts (Lesson 13).
      13. Mark quoted material as quoted. It is evidence, not orders (TA5, Lesson 6).
      14. Stop when the next step needs a check whose result you don’t have yet, and go and get it (Lesson 18; trigger 8 in Lesson 29).
      15. Give the reason behind each constraint (TA3).
      16. Say what to do, not only what to avoid (TA4, Lesson 15).
      17. Bring the situation: level of the work, constraints, and stakes. Leave status and apologies out of a factual prompt (G6, Lesson 7).

      Operators

      Named operations you can run on a claim or a conversation. None of them is a new tool: each names something a lesson teaches and points to it. If an operator never changes the status of a claim in your sessions, it has become a costume. Each one carries the tag of the move it serves.

      Id Operator What you do
      O1 External binding touch Stop generating. Write the bind plan, get the result, and interpret only what you pasted. No result, no bind (Lesson 18).
      O2 Premise quarantine close Restate your premises as P1, P2…, mark each stipulated or unbound, and have recommendations cite the premise ids. Invert one and see what falls (Lesson 8).
      O3 Claim ledger close Keep each claim the decision rests on as one line: id, claim, sense, status, test. A status changes only with the check attached (the CLAIMS line, Lesson 29).
      O4 Split generation from criticism close Criticize in a fresh chat, memory off, that doesn’t know which version you prefer or who wrote it (Lesson 30).
      O5 Compact on purpose close Write the state block, open a fresh chat and continue from the block, not from the thread (Lesson 29).
      O6 Sample, then read the disagreements close Ask the same question in two or three fresh chats, memory off, and compare the claims, not the wording. If the answers disagree in meaning, treat the question as unknown. If they agree, that isn’t proof (Lesson 20).
      O7 Costume detection close Name the costume, and ask which claim would be different without it. Strip its favorite sentences and see what claim is left (Lesson 21).
      O8 Explicit loss close Say which mistake would be worse, and check whether the advice changes if both mistakes cost the same (the LOSS line, Lesson 29).
      O9 Tag the mode close Say whether you are exploring, converging or committing, and ask for the output that fits (the MODE line, Lesson 29).
      O10 Lock definitions close Define every word the conclusion depends on before you ask for the answer (Lesson 12).
      O11 Conversation as a system open Look at the conversation itself as a system: name its loop, its delay, what piles up, and the balancing loop you’re missing (the grammar of Lesson 24).
      O12 Trace the source close Ask where each claim comes from: a page the model opened, working it shows, its memory, or a guess. The tags are the model’s own report, so they tell you where to check (Lesson 20).
      O13 Ask for the distribution open Ask for several substantively different answers with rough probabilities, unlikely ones included (Lesson 27).
      O14 Polyphony open Convene voices that answer one another, and keep their disagreements instead of averaging them (Lesson 27).
      O15 Attribution swap close Present your work as someone else’s, or the rival’s as yours, and compare the critiques (Lesson 8).
      O16 Clean the prompt close Have the model rewrite your question without your opinions and emotions, then ask the cleaned version in a fresh chat, memory off (Lesson 8).
      O17 Change the representation open Ask for the problem as a drawing, a table or a timeline (Lesson 27).
      O18 Reverse the roles touch You explain it in your own words, with the chat closed. Where you get stuck is the result. The model can then look for gaps in what you wrote, and each gap it names is a claim to check (Lessons 4 and 5).
      O19 Language swap close For a factual question, ask again in another language you read, as an invariance check (Lesson 21).
      O20 Make abstention usable close Welcome “unknown”, and treat a claim you’ll act on as unanswered until it has a source you can open or working you can recompute, whatever tag it carries (Lesson 20).

      Failure modes and how they feel

      Each failure is described by how it feels from the inside, because mid-session you won’t be reading a list. You’ll be having the feeling and calling it progress.

      How a failure feels

      Failure It feels like Interrupt
      Fluent captivity You learned a lot, and can’t restate one testable claim Write the claims your decision rests on as one-line statements, each marked bound or unbound, with no prose (the CLAIMS line, Lesson 29)
      Sycophantic lock Being unusually well understood The swap test: the same facts in a fresh chat, memory off, with the opposite preference stated. Then label your beliefs P1, P2… as hypotheses and have them evaluated before they are used (Lesson 8)
      Framework costume Sophistication; nothing was excluded Strike the framework’s vocabulary from the answer. Is there still a claim you could check? (Lesson 21)
      Precision mirage “Finally, it’s concrete” Ask how the number was measured. If there is no procedure, treat it as a rough range and mark it unbound (Lesson 18)
      Horizon escape Finally talking about the real issue (civilization, psychology…) Ask which level you are on, then go back to the pinned question and say what changed (Lesson 25)
      Local-optimum tutoring A productive lesson, and no decision Ask whether the next concept would change what you do. If you can’t tie it to DONE WHEN in one sentence, park it and decide (Lesson 29)
      Premise laundering The mess has been organized, so it must be understood Find the premise the tidy picture rests on, suppose the opposite, and see what still stands (Lesson 8)
      Critique theater Fairness; no claim changed status Hand the work to a fresh chat, memory off, to attack one named assumption or build a rival plan, and do the rewrite yourself. A devil’s advocate in the same thread is a costume (Lesson 30)
      Context rot Smarter and more confident, and contradicting turn three Write the state block and start a fresh chat with it (Lesson 29). Leave out the rejected tries, and put the passage the answer depends on next to the question (Lesson 3)
      Agreement cascade Momentum, alignment, speed Hand the plan to a fresh chat, memory off, with the adversarial packet: find the single most damaging flaw, and “broadly sound” is not an answer. Then state the option you rejected at its strongest, as its best defender would (Lesson 30)
      Emotion laundering Started angry or afraid, now feel rational Split it: what happened / what I feel / what I want to happen (Lesson 22). Which action still stands on what happened alone?
      Tool cosplay Due diligence Paste the result itself. A plan to look something up is not a lookup, and without a result the claim stays unbound (Lesson 18)
      Holon collapse Seeing the whole picture, because many things were mentioned Fill in the holon card and name the one level in focus this turn (Lesson 25)
      Derailment Being a serious thinker while the goal stalls Name the trigger from the list in Lesson 29, or drop the challenge and return to the goal
      Silent compliance Excellent service, real craft, wrong goal Before an expensive commit, check triggers 2 and 5 anyway: Contradiction and Proxy capture (Lesson 29)
      Protocol costume Meta-confidence Point at one claim whose status changed. If there is none, drop the framework vocabulary for three turns (Lesson 21)

      Further failures

      Panorama. Mechanism: “consider everything” plus holistic vocabulary makes every connection look relevant, with no boundary and no decision. Feeling: depth, even awe. Everything is connected, and no decision is closer. Interrupt: what is the relation, and where does it stop holding? Another consideration is not a relation (Lesson 23).

      Reframe laundering. Mechanism: a reframe is a cheap, impressive continuation, and the new frame quietly replaces the goal without a trigger or a brake. Feeling: insight. “The real question is…” The original done-condition no longer applies, and nobody decided that. Interrupt: name trigger 4 (Framing), write the action under both frames, and choose keep, revise or split (Lesson 29).

      Mitigated correction. Mechanism: politeness norms in the training data, and in you, soften disagreement into hedges and subordinate clauses. Feeling: it agrees with you, with nuance. The correction was there, after the “that said.” Interrupt: “List every point where you disagree with me, as bare statements, first” (Lesson 14).

      Persona inflation. Mechanism: prestige does not add the fact. Irrelevant biography can move accuracy down. Feeling: authority. Interrupt: drop the costume; give the constraints an expert would have (Lessons 7 and 15).

      Crutch learning. Mechanism: practice with help and performance without it come apart. Help that does the thinking for you lifts your score while you have it, and lowers it once the help is gone. evidence Feeling: ease, “I get it.” Interrupt: do the load-bearing part yourself, and check yourself later without help (Lesson 4).

      Comprehension debt. Mechanism: work piles up faster than anyone’s understanding of it. This happens with code, and also with a contract you signed after the model summarized it, or a plan you approved without reading. Feeling: speed, then fear of touching it. Interrupt: read what you would have to defend. Ask for an explanation of the existing thing before any change (TC13, TC14).

      Log-free success. Mechanism: “done” in a summary is a claim like any other, and the model may have written both the work and the check. Feeling: done. Interrupt: look at the output itself: the confirmation email, the sent folder, the test log. If the model could edit the check, the check is not a pass (Lesson 18).

      Citation mirage. Mechanism: citation formats are high-probability shapes, and a live link to a relevant page can still fail to support the sentence. A list of sources shows what the search returned, which is different from what was read. Feeling: rigor, because there are references. Interrupt: open the passage, not only the link (Lesson 19).

      Explanation compliance. The reason arrived, so the question feels settled. Interrupt: open the source, or write one way the answer is wrong, before reading the why (Lesson 8).

      Memory dominant. Feeling: it already knows you. Interrupt: read the app’s memory page, and delete conclusions it drew about you. If you didn’t say it today, treat it as unverified (Lesson 2).

      Middle way. It feels like wisdom, and nothing was negated. Interrupt: each original claim must still be visible, changed, and the new sentence must have been unavailable without them. Otherwise retract one (Lesson 26).

      Vigil. It feels like due diligence, line after line. Interrupt: compare the log to a written expected state. Do not wait to notice an unspecified anomaly (Lesson 29).

      Action handoff. It feels like velocity. The plan was accepted by being continued. Interrupt: you name the action. The model may draft it. If you cannot say what you would have rejected, you are out of the decision (Lesson 4).

      Back matter

      Glossary and sources

      Glossary

      Each term in a line or two, with the lesson where it is earned.

      Adversarial pass
      A critique from outside your thread: a fresh chat with memory off, a named standard, the author hidden. It shows what to check and is not itself the check (Lesson 30).
      Agent
      A model with tools that acts for you: it searches, opens files, runs code, fills in forms and works through steps without you (Lesson 3, Lesson 6).
      Assistant
      A base model after training on human preferences. Its tone, its refusals and its pull toward the typical, agreeable answer come from that training (Lesson 3).
      Base model
      A model that only continues a document, with no assistant persona and no trained refusals. Most people never talk to one (Lesson 3).
      Bears load
      A claim bears load if your decision would change were it false (Lesson 18).
      Bind, binding
      Tying a claim to a check that isn’t another piece of model writing: a source you opened, a number you recomputed, a test you ran, a person who was there. A claim is unbound until it has been put to such a check and bound after it. If the check goes against it, it is contradicted (Lesson 18).
      Borrowed senses
      What a model has when it uses search, code, files or a browser. The thing to check is the trace of that touch (the page, the file, the output), not the model’s paragraph about it (Lesson 3).
      Boundary marker
      A contradiction where both sides are true in different contexts. Naming the boundary turns it into a map (Lesson 26).
      Brake
      A challenge to the goal that names its trigger, shows the action under both goals and waits for your answer: keep, revise or split. With an agent that acts, the brake is the product’s approval step (Lesson 29).
      Center
      What your picture of a thing is organized around right now: what you currently take it to be. A working choice, not the thing’s essence, purpose or main part (Lesson 23). The center of a conversation is the question you are actually asking and the frame you are asking it in (Lesson 29).
      Ceremony
      How much discipline a conversation gets, set by the stakes and by how hard the decision is to undo. There are three levels: 0 the floor, 1 the Six, 2 full discipline (Lesson 29).
      Charity
      Reading a source, a person or a model at its strongest before you judge it. The audit comes after (Lesson 9).
      Claim-status tags
      Labels you ask the model to put on each claim: opened, worked out, from memory, guess. They show where to look and are not a check (Lesson 20).
      Close
      The move that narrows: define, choose, commit, write down the state (Lesson 5).
      Compaction
      In a long session, the tool replaces earlier turns with a summary to make room. Details can drop out, including a constraint you stated early (Lesson 3).
      Consilience
      Independent lines of evidence, each with different blind spots, converging on one conclusion (Lesson 17).
      Context
      Everything the model has in front of it when it answers: the conversation, pasted files, the app’s instructions, and memory if it is on (Lesson 2).
      Costume
      A surface that looks like rigor, neutrality, criticism or depth without changing the status of any claim that bears load (Lesson 21).
      Cynefin
      Say “kuh-NEV-in”. A framework that sorts situations by how cause and effect relate: clear, complicated, complex, chaotic, confused (Lesson 28).
      Delta
      What changed since the pinned question, said on return. “Nothing changed” counts (Lesson 25).
      Derailment
      A sound goal keeps being challenged because the task got hard or a more interesting question turned up (Lesson 29).
      Deserved interlocutor
      Ukhtomsky’s phrase for the person who has done the inner work to meet the other as it is, and so can open without being fooled and close without shutting down (Lesson 7).
      Design tension
      A contradiction where you want both sides and they limit each other. Separate them, or choose and record what you gave up (Lesson 26).
      Dominant
      Ukhtomsky’s word for a focus of attention that pulls everything you notice toward itself. It helps when it is a question you are holding, and blinds when it is a conclusion you have already reached (Lesson 4, Lesson 10).
      Double
      Ukhtomsky’s word for the self you keep meeting in others when your attention rests on your own wishes, fears and rivalries. With a model, it is your hope coming back as a finding, or a person or a machine that you project onto it (Lesson 8).
      Drift
      A turn that quietly changes the goal, widens the scope or answers a neighboring question (Lesson 29).
      Effort level
      The setting some products give you for how much a reasoning model works before it answers. It pays on hard, checkable work and does little for taste or a quick lookup (Lesson 3).
      Fact map
      For a tangled situation: the facts collected on a page before any explanation, sorted by type, with where to check each, what is missing, and arrows you draw yourself (Lesson 10).
      Fertile tension
      A hedge that sits on a real fork in the question and is worth probing before you close. The test: ask what fact about your case would settle it (Lesson 14).
      Floor
      The few habits every conversation deserves: the eight on the Start here page. It is ceremony level 0 (Lesson 29). The appendix keeps a longer prompting checklist behind them, under Prompt basics.
      Foothold
      A temporary stable point in an open inquiry: something that holds for now and that the next question can be built on (Lesson 27).
      Frame
      The picture a wording calls up: what kind of situation this is and what belongs in it (Lesson 15).
      Fresh context
      A new chat with no history and memory off, or a temporary chat set to use nothing the app knows about you. It is the instrument for critique, swap tests and second estimates. Also called a fresh chat or a clean chat (Lesson 2, Lesson 8).
      Genre
      A named kind of document whose conventions, including its standard of evidence, come with the name: a referee report, a postmortem, a buyer’s guide (Lesson 15).
      Grounding
      Evidence that the other side understood. With a model you have to ask for it, and the cheapest form is a restatement before the work starts (Lesson 13).
      Hinge
      The single feature that, if it were different, would make this a different situation. Under time pressure it is the one thing to check in the source itself (Lesson 29).
      Holon
      Anything seen as both a whole made of parts and a part of a larger whole. It is the reason every conversation about a system sits at some level (Lesson 25).
      Invariance battery
      A set of one-minute checks that change the lean, the wording, the language, the format, the order, the role or the model, to see whether an answer holds (Lesson 21).
      Invariant-so-far
      An answer that survived changes of wording, frame and asker. It is still not checked against the world (Lesson 21).
      Jagged abilities
      Strong on one task and weak on a neighboring one, with human difficulty a poor guide to which is which (Lesson 3).
      Ladder of abstraction
      The climb from one concrete case to wider and wider categories. Move up and down it on purpose (Lesson 16).
      Ladder of inference
      The steps from observable data through selected data, meanings, assumptions and conclusions to beliefs and actions. When you disagree, walk back down to the data (Lesson 16).
      Library
      The text a model was trained on, seen as a compressed library of human reports. It is weighted by how often things were said, which is different from how true or important they are (Lesson 1, Lesson 19).
      Lock, unlock
      To lock a word is to fix the meaning of a word your conclusion depends on. To unlock it is to ask for its other meanings when you are exploring (Lesson 12).
      Log
      The record of what actually happened: the confirmation email, the sent folder, the tool’s output, the test run. It is what you read when a model reports that something is done (Lesson 18).
      Mode
      The single most typical answer to a request, which is what you get when you ask for “an idea” (Lesson 27). The MODE line of a state block is a different thing: it names the kind of work the session is doing (Lesson 29).
      Mode of existence
      A thing’s own logic of working, which shouldn’t be scored by another thing’s criteria (Lesson 3, Lesson 23).
      Moves, the three
      Open, close and touch: the three steps of the one rhythm the book teaches (Lesson 5).
      Name the act
      Saying which act you want (explore, explain, evaluate, decide, draft, build, check or teach) before the model picks one for you (Lesson 13).
      One-way door, two-way door
      A two-way door is a decision you can walk back: rewording an email, swapping a recipe. A one-way door is one you can’t, or only at great cost: quitting a job, signing a lease. Two-way doors get the floor, and one-way doors get full discipline (Lesson 29).
      Open
      The move that widens: other reports, other meanings, other options, a contradiction held (Lesson 5).
      Outer log
      The line of the state block that records each challenge to the goal, the trigger named, and whether you answered keep, revise or split (Lesson 29).
      Panorama
      Everything relates to everything, and nothing is decided. The costume of open thinking (Lesson 5, Lesson 23).
      Parked
      A good side idea written down as not this conversation’s job, and not discussed now. If you can name a trigger for it, it goes through the brake instead (Lesson 29).
      Polarity
      Two poles that depend on each other and never settle. You map and manage them instead of solving them (Lesson 26).
      Polyphony
      Independent voices that answer each other, with their disagreements kept and no narrator to reconcile them (Lesson 27).
      Portal
      Using the model to find terms, camps and kinds of sources, then going through to the sources themselves (Lesson 19).
      Premise laundering
      A guess of yours that comes back later in the thread as an established fact (Lesson 8).
      Probe
      A small trial you can afford to lose, with one signal for doing more of it and one for stopping. It is the way to touch a complex situation (Lesson 28).
      Product memory
      What an app stores about you between chats and can read into a new one. It is part of the context, and you didn’t type it there (Lesson 2, Lesson 6).
      Proxy capture
      A measure rising while the aim it stood for does not. It is also trigger 5 for challenging a goal (Lesson 24, Lesson 29).
      Quiet part
      What is true and common but rarely written about with feeling: the base rate, the dull majority case, the people who know and don’t post (Lesson 19).
      Reasoning model
      A model that spends extra computation on the problem before it answers and hides most of that work. What a product shows as “thinking” is at most a summary (Lesson 3).
      Remembered source, opened source
      A source the model recalls from training, and a page it fetched in this conversation. Both can be cited the same way. Only the second was read, and neither is checked until you open it (Lesson 19).
      Role with a standard
      A role that says what the answer must contain. A role that only changes how the answer sounds is a costume (Lesson 15, Lesson 21).
      Sample
      One more answer from a model. Two samples agreeing is not a check (Lesson 18).
      Session Card
      The form that holds a state block and writes the prompts that open, continue and close a session (Lesson 29).
      Signpost
      An observable sign with a threshold and a date that tells you an assumption has failed (Lesson 30).
      Silent compliance
      The goal is wrong and the model helps anyway, without saying so (Lesson 29).
      Six, the
      The six habits added to the floor when the outcome matters. Ceremony level 1 (Lesson 29).
      State block
      A short note you own that says where the conversation stands: the goal with its use and done-condition, your premises, the claims and whether each is checked, your commitments, and what is parked. You paste it into a long thread or a fresh one (Lesson 29).
      Status
      Where a claim stands right now: stipulated (taken as given for this conversation), unbound (not checked yet), invariant-so-far (it survived a change of wording or framing) or bound (checked against something outside the conversation). A bound claim that the check went against is contradicted (Lesson 17).
      Steelman
      Stating a view or a plan at its strongest, as its best defender would, before you criticize it. It is the principle of charity used as a first step (Lesson 9, Lesson 30).
      Stipulated
      Taken as given for this conversation and not checked. It is one of the statuses a claim or a premise can have (Lesson 17).
      Stock, flow, feedback loop
      A stock is something that accumulates: skill, trust, savings, lost sleep. Flows fill and drain it, delays separate causes from effects, and a feedback loop is either reinforcing, where more leads to more, or balancing, where more leads to less (Lesson 24).
      Swap test
      Ask the same question in a fresh chat, memory off, while stating the opposite preference. If the answer flips with no new evidence, it was tracking you (Lesson 8).
      Teach-back
      Explaining what you learned in your own words with the chat closed (Lesson 2). It is a touch turned inward (Lesson 5). Asking the model afterwards to find the holes in what you wrote is a critique, which shows you where to look (Lesson 4).
      The third
      What arises between two sides and belongs to neither: in a conversation, the shared sense of where things stand, or an idea neither brought. With a model its visible form is the growing text of the conversation (Lesson 2). The same question can be asked of any two things (Lesson 23).
      Three Others
      The model, people and reality: the three parties in every serious conversation. Most failures come from treating one as another (Lesson 9).
      Touch
      The move that checks against something that can say no. Outward: a source opened, a number recomputed, a test run, a person asked. Inward: you explain or apply it with the model out of the room. Another model answer is never a touch (Lesson 5).
      Trigger
      One of eight named reasons that make a challenge to the goal legitimate: harm, contradiction, false premise, framing, proxy capture, wrong level, silent override, check first (Lesson 29).
      Typical case
      What the model fills in for anything you didn’t say: the most common asker, situation and answer (Lesson 2, Lesson 7).
      Unanswered
      A claim you would act on that comes without a source you can open or working you can recompute, however confident it sounds (Lesson 20).
      Well
      A discipline or habit of mind that sees one part deeply and loses sight of the rest (Lesson 1).
      Wholeness
      The web of relations that makes a thing itself. A whole is a thing already cut out of the world, and wholeness is what the cut can lose (Lesson 23).

      Sources

      Entries are grouped by topic and run alphabetically inside each group. Links are given only where they were checked. Everything else is cited by author, title and year, so you can find it and open it yourself, which is what this book asks you to do with any source. A preprint has not been through peer review. A vendor document is a company describing its own product. An industry measurement was published by a company, not a journal. Pages that change carry the date they were read.

      Foundations

      • Andrey Kurpatov (Андрей Курпатов), «Заслуженный собеседник. Искусство диалога с искусственным интеллектом» (2025). The title in English: The Deserved Interlocutor: The Art of Dialogue with Artificial Intelligence.
      • A. A. Ukhtomsky (А. А. Ухтомский), «Заслуженный собеседник» (1997), a collection of his letters, diaries and notes, cited by page.
      • A. A. Ukhtomsky (1923). «Доминанта как рабочий принцип нервных центров» (“The dominant as a working principle of nerve centers”).
      • John Godfrey Saxe (1872). “The Blind Men and the Elephant.”
      • The older versions of the parable: the Buddhist Udāna 6.4; the Jain teaching of anekāntavāda (many-sidedness) with its qualifier syāt; Rumi, Masnavi, “The Elephant in the Dark” (Nicholson’s translation).
      • Mike Caulfield (6 November 2025). Post on his “Get it in, track it down, follow up” framework, the source of “a portal, not a portrait” and of the line on knowledge maps. LinkedIn.

      How language models behave

      Agreement and sycophancy

      Confidence, calibration and abstention

      Models judging and debating models

      Models and products: vendor documents, evaluations and reports

      People working with automation and AI

      Thinking, attention and judgment

      Learning

      Language and conversation

      • Austin (1962). How to Do Things with Words.
      • Bakhtin (1963; English 1984). Problems of Dostoevsky’s Poetics.
      • Besco (1995). Releasing the hook on the copilot’s catch 22. Proceedings of the Human Factors and Ergonomics Society 39th Annual Meeting. The P.A.C.E. steps: probing, alerting, challenging, emergency warning. First presented in 1994 at the 25th International Seminar of the International Society of Air Safety Investigators (“To intervene or not to intervene? The copilot’s catch 22”).
      • Brown & Levinson (1987). Politeness: Some Universals in Language Usage. Cambridge University Press.
      • Buber (1923). Ich und Du (in English, I and Thou).
      • Clark & Brennan (1991). Grounding in communication. In Resnick, Levine & Teasley (eds.), Perspectives on Socially Shared Cognition. American Psychological Association.
      • Dennett (1991). Consciousness Explained. The “multiple drafts” model.
      • Dennett (2013). Intuition Pumps and Other Tools for Thinking. The source of Rapoport’s rules.
      • Grice (1975). Logic and conversation. In Cole & Morgan (eds.), Syntax and Semantics 3: Speech Acts.
      • Hayakawa (1941). Language in Action.
      • Helmreich, Merritt & Wilhelm (1999). The evolution of Crew Resource Management training in commercial aviation. International Journal of Aviation Psychology 9(1). The history of the training that airlines adopted after the Tenerife (1977) and United 173 (1978) accidents.
      • Johnson, W. (1946). People in Quandaries.
      • Klein, Moon & Hoffman (2006). Making sense of sensemaking 1: Alternative perspectives; and 2: A macrocognitive model. IEEE Intelligent Systems. Two essays.
      • Korzybski (1933). Science and Sanity.
      • Lakoff (2004). Don’t Think of an Elephant!
      • Pomerantz (1984). Agreeing and disagreeing with assessments: Some features of preferred/dispreferred turn shapes. In Atkinson & Heritage (eds.), Structures of Social Action. Cambridge University Press.
      • Sacks (1987). On the preferences for agreement and contiguity in sequences in conversation. In Button & Lee (eds.), Talk and Social Organisation.
      • Searle (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.
      • Senge, Kleiner, Roberts, Ross & Smith (1994). The Fifth Discipline Fieldbook. The popular source for Chris Argyris’s ladder of inference.
      • Steen, Reijnierse & Burgers (2014). When do natural language metaphors influence reasoning? PLoS ONE.
      • Thibodeau & Boroditsky (2011). Metaphors we think with: the role of metaphor in reasoning. PLoS ONE.
      • Thibodeau & Boroditsky (2015). Measuring effects of metaphor in a dynamic opinion landscape. PLoS ONE.
      • Walton & Krabbe (1995). Commitment in Dialogue.
      • Wolff & Holmes (2011). Linguistic relativity. WIREs Cognitive Science 2(3).
      • Yudkowsky (undated). Crocker’s Rules. sl4.org, the site of the SL4 mailing list. Named for Lee Daniel Crocker. Read 10 November 2026.

      Truth, evidence and decisions

      • Atkin, Schoar & Shinde (2023). Working from Home, Worker Sorting and Development. NBER Working Paper 31515.
      • Bloom, Liang, Roberts & Ying (2015). Does working from home work? Evidence from a Chinese experiment. Quarterly Journal of Economics 130(1).
      • Bloom, Han & Liang (2024). Hybrid working from home improves retention without damaging performance. Nature.
      • Christensen (2010). Higher-order evidence. Philosophy and Phenomenological Research.
      • Collins (2010). Tacit and Explicit Knowledge.
      • Dewar, Builder, Hix & Levin (1993). Assumption-Based Planning. RAND.
      • Elga (2007). Reflection and disagreement. Noûs.
      • Ellsberg (1961). Risk, Ambiguity, and the Savage Axioms. Quarterly Journal of Economics.
      • Flyvbjerg (2006). From Nobel Prize to Project Management: Getting Risks Right. Project Management Journal.
      • Hardwig (1985). Epistemic Dependence. Journal of Philosophy.
      • Hauswald (2025). Artificial Epistemic Authorities. Social Epistemology.
      • Hernán & Robins (2020). Causal Inference: What If.
      • Heuer (1999). Psychology of Intelligence Analysis. CIA Center for the Study of Intelligence.
      • Howard (1966). Information Value Theory. IEEE Transactions on Systems Science and Cybernetics.
      • Kelly (2005). The epistemic significance of disagreement. Oxford Studies in Epistemology 1.
      • Klein, Calderwood & MacGregor (1989). Critical decision method for eliciting knowledge. IEEE Transactions on Systems, Man, and Cybernetics.
      • Klein, G. (2007). Performing a project premortem. Harvard Business Review.
      • Knight (1921). Risk, Uncertainty and Profit.
      • Lakatos (1970). Falsification and the methodology of scientific research programmes. In Lakatos & Musgrave (eds.), Criticism and the Growth of Knowledge.
      • Lipton (2004). Inference to the Best Explanation (2nd ed.). Routledge.
      • Lovallo & Kahneman (2003). Delusions of Success. Harvard Business Review.
      • Mitchell, Russo & Pennington (1989). Back to the future: Temporal perspective in the explanation of events. Journal of Behavioral Decision Making 2.
      • Pearl & Mackenzie (2018). The Book of Why.
      • Schweiger, Sandberg & Ragan (1986). Group approaches for improving strategic decision making: A comparative analysis of dialectical inquiry, devil’s advocacy, and consensus. Academy of Management Journal 29(1).
      • Schweiger, Sandberg & Rechner (1989). Experiential effects of dialectical inquiry, devil’s advocacy and consensus approaches to strategic decision making. Academy of Management Journal 32(4).
      • van Fraassen (1989). Laws and Symmetry. Oxford University Press.
      • Vane (1971). Inhibition of prostaglandin synthesis as a mechanism of action for aspirin-like drugs. Nature New Biology. See the Nobel Prize page for 1982.
      • Wack (1985). Scenarios: uncharted waters ahead; and Scenarios: shooting the rapids. Harvard Business Review. Two articles.
      • Whewell (1840). The Philosophy of the Inductive Sciences.
      • Wilson (1998). Consilience: The Unity of Knowledge.

      Finding information: search, citations and sources

      Systems, problems and new ideas

      Software and agents

      Law, official guidance and court rulings