That sounds like a small distinction. It is the whole difference between a note you can trust next month and a note you have to re-check.

Why a rewrite is the wrong default​

An AI asked to "clean up" a note will happily do more than clean. "Maybe the cover is too dark?" becomes "The cover is too dark." Three fragments about a supplier become a tidy paragraph with a recommendation at the end that nobody made. Each change reads well. Together they turn your record of what happened into the model's guess about what probably happened.

The damage shows up later. A note is only useful if you can act on it without re-checking, and you cannot tell, a month on, which sentences were yours. So we drew the line where a good secretary draws it: take the words down properly, put them in the right file, and keep your opinions out of it.

What it fixes, every time​

The first half of the job is the scribe's. Every quick note gets the same treatment, however short:

• Wording. Typos, a stray letter, the word speech recognition misheard, grammar and word order are corrected — in a one-line note too. An earlier version left short notes alone; the typos that survived were the reason to stop.
• Form. Filler and false starts go; when you correct yourself mid-sentence, only the corrected version stays; the point comes first; something you have to do becomes a task. A short line does not get headings or bullets it does not need.
• Place. The thought goes where you would have put it: under its heading, after the item it continues, in the list if that section is a list.

If you want a line kept exactly as you typed it, say so, and only the punctuation is touched.

What it will not do​

The second rule is the one that makes the first safe. Every point, name, number and hedge you gave stays. A question stays a question — it is not answered for you. "Maybe" does not become "definitely". Nothing is added: no examples you did not give, no summary at the end, no next step you did not decide.

That includes requests for new writing that arrive in the middle of a quick note. "Turn the above into an email to the client" is real work, and it is never slipped into the note as if you had written it. It is kept as a task, with an offer to run it — or, if you meant it as an instruction from the start, you press ⌘Enter instead of Enter and it is done as its own change, previewed word by word. Edits to what is already there — rename this heading, merge those two items, move that line — are carried out directly, because they rearrange your words rather than inventing new ones.

When it is not sure, it says so​

Correcting spelling means guessing, and guesses about names are the ones that go wrong. So the corrections it is unsure of are listed — "lina → Lena" — beside the change, before you accept it. Most of the time the list is empty. When it is not, it is usually the one word that would have been embarrassing in an email.

And nothing is written until you accept. A quick note arrives as a marked change on the note: the corrected text, where it is going, and an Accept and a Reject. Parts of the note it is not filing into are not touched at all.

Tidying the whole note, only when you ask​

A week of quick notes piles up in the order things occurred to you. A Tidy whole note button reorganises the lot: it fixes typos across the note, brings related points together, keeps a point you wrote twice once, and adds short labels when the note covers several topics. Dated entries keep their order. The same rule holds — the shape changes, the content does not — and every change is a revision you can reject.

It runs only when you press it. A note that reorganised itself every time you added a line would never look the way you left it, and a note you cannot find your way around is worse than a messy one.

What it costs​

The rule has a price, and it is better to know it up front.

• The note stays your note. If you jot fragments, you get corrected, filed fragments — not prose. A tool that writes for you will produce nicer-looking pages; they will just be less yours.
• Some fixes need your eye. Names and in-house jargon are where corrections go wrong. The uncertainty list catches most of it; it cannot catch a confident mistake on a name it has never seen.
• New writing is a separate step. Turning a log into an update is one key away, but it is a step you take on purpose, and its draft deserves a read before it goes out.

What we are not claiming​

We are not claiming it never gets a word wrong — it does, which is why changes wait for you. We are not claiming it is a task manager, a calendar or a meeting recorder: "remind me" becomes a to-do in the note, not an alert. And we are not claiming fixing is neutral: correcting a word is a judgement. The promise is narrower and checkable — it does not add what you did not say, and it shows you every change it makes.

How it works in the app: AI quick notes, with the keys and the Tidy whole note button in voice and quick notes, step by step. A week of it in practice: keep a project log that writes your weekly update.

That sounds like a small detail. It is the whole difference between an edit you can send to a client and an afternoon spent rebuilding a document that used to be fine.

Why pasting a Word file into a chat loses the document​

A chat assistant sees your document as plain text, and it answers in plain text. Everything that made the file a Word file — the company template, the heading styles the table of contents is built from, the automatic numbering, the header with the logo, the comments a colleague left — is gone on the way in, and does not come back on the way out.

So the round usually goes like this:

1. Copy the document into the chat, or upload it, and explain what you want.
2. Get back a new version, with changes you asked for and a few you did not.
3. Compare it with the original by eye to find what moved.
4. Paste the good parts back into Word and repair the styles, spacing and numbering the paste broke.
5. Send it on with no record of what changed — or turn on Track Changes and retype the edits by hand, so your reader can see them.

Step 5 is the one people forget. Whoever receives a contract, a proposal or a thesis chapter in Word expects to review it: to see each change, accept the ones they agree with and reject the rest. A clean file with silent edits takes that away from them.

What “AI edits with tracked changes” should mean​

Three tests separate a real answer from a workaround:

• The same file. The .docx in your folder is the one that changes. No export, no paste-back, no second copy to keep in sync.
• Only the words change. A revised paragraph keeps its style; the heading is still a heading; 2.1 is still numbered 2.1. Nothing outside the edited text is rebuilt.
• Every edit is a revision mark. Open the file in Word and each change is a tracked change with an author, which the next reader can accept or reject in the Review tab they already know.

And one test before any of that: you see every edit before it touches the file. An AI that writes into your original first and lets you undo afterwards has made you the proofreader of its guesses.

How to do it in MindMax​

1. Open the folder the document lives in. MindMax works on the folder as it is; the .docx is not imported or converted.
2. @ the Word file in the assistant and say the change once. Write it the way you would to an editor: “cut the executive summary to half a page and use the client’s term, platform, throughout”, or “answer the reviewer’s three points in section 4”. A compound request is fine; it is split into separate edits.
3. Read what it intends. The revision opens with a short summary of what it is about to change, then groups of edits — each with a title and a one-line reason — marked on the document’s own text.
4. Keep what you agree with. Accept or reject group by group, or all at once. Reject everything and the file is exactly as it was.
5. Apply. At the bottom of the review, Write as Word tracked changes is ticked by default. Leave it, and every accepted edit becomes a revision mark in the same .docx. Untick it if you want a clean edit instead.
6. Send it. The tracked changes carry your name — from your MindMax account or your computer’s user name, or one you set in Settings — so the reader sees them as yours, because they are: you chose them.

In MindMax this is the same review every AI edit goes through, whether the document is a Markdown note or a Word file. The only difference is where the accepted edits are written.

What happens to the formatting​

It is left alone. Only the text of the paragraphs that change is rewritten; everything else in the file is carried over as it was.

• Styles and fonts. A changed paragraph keeps its paragraph style, and new words take the formatting of the words they replace.
• Headings and numbering. Heading styles and automatic list numbering stay Word’s own.
• Headers, footers and footnotes. Untouched — they are outside what the AI revises.
• Images, fields and comments. Pictures, page numbers, the table of contents, cross-references and the comments your colleagues left stay where they were.
• Tables. Text in cells can be revised; the table’s columns and merged cells are not reshaped.

Before anything is written, the edited file is read back and checked against the edits you accepted. If the check fails, nothing is written and the original stays as it was. If the document is open in Word, you are asked to close it first. If the file changed after the edits were suggested, you are offered fresh ones instead of stale edits being forced in. And the version before the revision is kept in the file’s history in MindMax, so a change you regret is one step back.

When the document already has someone else’s tracked changes​

Documents that go back and forth collect tracked changes. If an AI adds its own on top, nobody can tell old from new.

When you open a Word file with pending tracked changes in MindMax, a bar at the top says how many there are and whose, with Accept all, Reject all and Open in Word. When you ask for a revision, the assistant suggests accepting the old changes first, so its edits are not mixed in with them. If you would rather go through them one by one, open the file in Word — or keep them and revise anyway.

What about PDFs?​

A PDF is a printed page, not text that can be safely rewritten in place, so MindMax never changes the PDF itself. Ask for a revision the same way and review the edits on the PDF’s text; when you apply, a Markdown file with the same name is created next to the PDF, holding its text with the edits you kept. Export that to PDF or Word when it has to be a file again. It will be a clean document in MindMax’s layout, not the original’s columns and fonts.

A scanned PDF needs its text recognised first. For a long PDF or a scanned one, the assistant tells you what to expect before it starts.

Where it goes wrong​

• Vague requests get confident edits. “Make it better” produces a pass that changes a lot and improves little. Name the problem and where it is.
• Not everything in the file is revised. Headers, footers, footnotes and text boxes are kept but not edited. A change that would add a table column, or that falls inside a page-number field or a content control, is left out and listed rather than forced in.
• Word comments are not instructions. Comments left in the file are kept, but the AI does not read them as requests. Put the change in the conversation.
• Only .docx. An old .doc file needs saving as .docx in Word first.
• Word’s format, checked in Word. The tracked changes are written the way Word stores them. How WPS, Pages or Google Docs display them has not been checked.
• A PDF copy loses the layout. Columns, fonts and page design do not carry over to the Markdown copy, and images may not come along.

What is not claimed​

This does not make the AI right. It makes its edits visible, small enough to judge one group at a time, and easy to reject — which is what you need when the AI is wrong. You still decide what the document should say.

AI edits for Word documents · Reviewing and revising with AI · Download MindMax

The point of a review is not the opinion. It is an opinion the author can act on, on the sentence it is about, in the file they already work in.

Why a chat review takes all afternoon​

Asking a chat assistant to review a document is fast. What comes back is a list: the argument in section 3 is weak, the evidence for the main claim is thin, a few terms are undefined. Then the real work starts:

1. Find the paragraph each point is about — and wonder whether it read that paragraph at all.
2. Throw out the points that are generic, wrong, or about a problem the draft does not have.
3. Rewrite the rest so they sound like a colleague and not a chatbot.
4. Open Word and add each one as a comment, on the right words, one at a time.

Steps 1 and 4 are clerical. Steps 2 and 3 are the review. A good tool should take the first kind off your hands and leave you all of the second.

How to do it in MindMax​

1. Put the draft in your folder and open it. A .docx, or a Markdown or text file. If there is a checklist, reviewer guidelines or a brief the draft has to answer, save it in the same folder.

2. Say what the review is for. Choose AI → Review this document…. In the review settings, write a focus in a line — "check whether the cost estimate holds up and the rollout risks; ignore wording" — and, if you have one, pick the criteria document. The draft is checked against it point by point.

The focus decides where the comments concentrate, not what the AI is allowed to see. A serious problem outside it is still raised. That matters: the worst review is the one that answered the question it was asked and missed the hole in the middle of the argument.

3. Let it work. It reads the whole document, chapter by chapter, starting with the one you are reading. Each comment is written to the author and sits on the passage it is about, as one of five kinds:

• a flaw in the method, argument or design;
• an unsupported claim, naming what the evidence is missing;
• an alternative the author never ruled out;
• an unclear passage, where a reader cannot tell whether it is right;
• a strength worth keeping.

All but strengths are marked major — must be addressed before the conclusion stands — or minor. Most comments are minor; major has to be earned. A review memo follows: overall assessment, major issues, minor issues, strengths, and the questions the author should answer.

4. Adopt, reword, ignore. On each comment, Adopt opens it with the AI's wording in place, for you to edit before it becomes your comment. Ignore the ones you disagree with. Select any passage to add the points only you would make — the ones that come from knowing the project, the field, or the person.

5. Export the reviewed copy. Export reviewed Word copy saves <name> - reviewed.docx next to the original. Tick the comments to include; optionally put the review memo at the top as an overall comment, and prefix comments with [Major], [Minor] or [Strength] so the author knows where to start. The original file is not touched, and any comments already in it are kept.

The author opens a Word file and finds comments in the margin, signed with your name. They do not need MindMax, and they do not need to know how you worked.

Whose words reach the author​

Only yours. The comments in the reviewed copy are the ones you adopted, in the wording you kept, and the ones you wrote yourself. The AI's other drafts are not exported, and neither is anything you discussed with it while deciding. If you would not sign a comment, do not adopt it.

Where it goes wrong​

• Confidential manuscripts. A review sends the document's text to the model you use. Many journals and funders do not allow submissions to be put into AI tools. Check first; with a model running on your own computer, the text does not leave it.
• PDF submissions. A PDF can be reviewed in MindMax, but the reviewed Word copy is made only from .docx, Markdown and text files. Ask the author for the source.
• Comments, not rewrites. The copy carries comments. Suggested wording goes inside a comment; nothing is changed in the author's text.
• Comments already in the file. They are kept, but the AI does not read or answer them.
• Repeated text. If the quoted words appear more than once, or are too short to place with confidence, the comment is listed with the reason instead of pinned to a guess. Add it by hand.
• It does not know your project. It will raise things you would not have and miss things only an insider would see. That is what the adopt step is for.

What is not claimed​

This does not make the AI a reviewer. It makes the clerical half of reviewing — finding the sentence, drafting the first version of the comment, putting it in Word — cheap, so that your time goes to the half that needs you: deciding what is true and what is worth saying.

AI document review · Review a colleague's draft, step by step · Download MindMax

Every AI reading tool has to decide what happens when the material is longer than the model can handle well. That decision is invisible in the answer, and it determines whether the answer can be trusted.

Option 1: paste the whole thing in​

Context windows are now large enough to hold many books, so this sounds like the obvious fix. It has two problems.

Fitting is not the same as attending. Research on long contexts — the best-known paper is titled Lost in the Middle — found that models recall material at the start and end of a long input far more reliably than material in the middle. A book pasted in whole is mostly middle. The answer will be fluent, and it will lean on whatever was easiest to see.

It is expensive every time. Every question re-sends the book. Ten follow-ups cost ten books.

Option 2: chunk it and retrieve​

The standard pipeline cuts the book into fragments, finds the ones most similar to your question, and answers from those. For "what year did the author move to Paris?" this works well.

For questions about the book as a whole, it fails quietly. "Where does the author change position?" has no single fragment that matches it; the evidence is spread across chapters. "What does this book never address?" has nothing to match at all. And the answer gives no sign that the decisive passage was never retrieved.

Option 3: read it like a reader​

What a person does with a book they have not read is neither of the above. They look at the contents, pick the chapters the question is about, read those properly, and flip back when something does not add up.

That is the approach MindMax takes:

1. Structure first. The chapter map comes from the table of contents the book was published with — or, for PDFs without bookmarks, from the headings — so the book can be navigated by its own structure.
2. Whole sections, not fragments. Once a chapter is chosen, it is read in full, with its structure intact.
3. Go back if needed. If what was read does not settle the question, another chapter is opened.
4. Say what was read. Every answer lists the sections it actually read, so "the book does not say" and "it did not look" are never the same answer.
5. Quote it. Every claim carries a verbatim quote that opens on the passage in the book.

Documents short enough to read whole — most papers, reports and single chapters — are simply read whole. The navigation only starts when the material is bigger than the model can attend to well, and that line is drawn well below what the model technically accepts.

What this does not fix​

• Badly built books. If the chapter metadata is broken, or the whole book is one giant file, there is no structure to navigate. The book still reads; whole-book questions get harder.
• Questions that need every page. "Count every mention of X" is a search job, not a reading job — use search inside the book for that.
• Judgement. Reading the right chapters gets you an answer grounded in the book. Whether the book is right is still your call.

What is not claimed​

We have not published a benchmark comparing these approaches, and we are not going to quote one we have not run. The argument here is about failure modes: pasting in hides attention loss, retrieval hides misses, and reading by structure — with a list of what was read — makes its own gaps visible.

The concept in full is full-text-first answering; the why-no-vector-database side of it is in why there is no vector database in it. For books specifically, see the AI EPUB and Kindle reader.

That line is deliberate, and this post explains where it sits and why — because "free tier" in most software means "a demo that runs out on day two", and that is not what we mean.

What never costs anything​

• Opening and reading PDFs, EPUB and Kindle books, Word, PowerPoint, Excel, images and Markdown.
• Highlights, comments and your own notes, kept in your folder.
• Search across everything in the folder.
• Writing and editing in the Markdown editor, including version history.
• Exporting to Word, PDF or images — it happens on your computer.
• Your files themselves. Everything MindMax produces lives in your own folder and stays readable if you stop paying.

None of these touch a model, so there is nothing to charge for.

What spends credits​

Anything where a model does work for you:

• Asking about a document, a folder or a boxed region, and Explain on a selected passage.
• An AI reading pass over a paper or a book — the larger the document, the larger the job.
• Translation of a document or a whole book.
• Recognising a scanned PDF, which is why it only starts when you press the button on that document.
• Canvas skills, inline AI in the editor, revisions, and long-form drafting.
• Voice transcription, and searching the web from a conversation on plans that include it.

The pattern: short questions are small jobs, whole-book passes and translations are large ones. Try a large job on one chapter before committing a shelf.

Three ways to pay for the AI part​

1. The free plan includes a small weekly AI allowance. There is no trial countdown — the app does not stop working when the allowance is used; AI work pauses until the week rolls over.
2. A paid plan adds a larger weekly allowance and no keys to manage. If you run out, AI work pauses until the week rolls over or you move up a tier. Nothing is charged automatically.
3. Your own API key, one you already pay for with a provider (a chat subscription is not one). Every feature works, with no allowance and no account. You pay your provider directly, and nothing stops a job that runs long — so keep an eye on usage.

You can mix them: point some kinds of work at your own key and the rest at the subscription. Credits are spent only on what goes through the subscription.

How to keep costs predictable​

• Ask about the part you need. A boxed figure or a selected passage is a smaller job than the whole paper.
• Follow-ups are cheaper than they look. A document read whole is reused between turns rather than paid for again with each question.
• Recognise and translate on purpose. Both are deliberate buttons, never run silently over a folder.
• Watch the usage view. It shows credits and tokens by task, so a surprising week has an explanation.

Why the line is drawn here​

AI usage is a real cost to whoever pays for it; reading a file on your own computer is not. Charging for the second to fund the first would mean holding your own files hostage to a subscription. So the free part is the software, and the paid part is the compute — which is also why using your own key is a first-class option rather than a reduced mode (more on that).

Current plans and allowances are on the pricing page; prices are not repeated here so this post does not go out of date. Setting up your own key: use your own API key. Download MindMax — the app costs nothing to try.

The last post in Thinking models in practice, a short series on classic methods and how to run them on your own material.

Where it comes from​

Variants go by many names — weighted decision matrix, option matrix, grid analysis, Pugh matrix after the engineer Stuart Pugh, who used a version to compare design concepts against a baseline. They share one idea: make the comparison explicit and the same for every option.

How to build one​

1. List the real options, including "do nothing" or "keep the current one". Three to five is typical; more usually means you have not narrowed the question.
2. Pick criteria that would actually change your mind. Cost, time, risk, reversibility, fit with what you want. Drop criteria every option scores the same on — they add rows of noise.
3. Weight the criteria if they are not equally important. Decide the weights before scoring, or you will tune them toward the answer you already like.
4. Score each option on each criterion. A simple 1–5 scale is enough. Write one line of justification per score; it is where the thinking happens.
5. Look at the result, then distrust it. If the winner surprises you, find out why. If it does not, ask which single score is carrying it.

A worked example​

Decision: should a small software company drop its monthly plan before a spring launch?

The matrix leans toward keeping things as they are, narrowly ahead of a partial change. The useful part is what it exposes: "drop for new customers only" loses mainly on simplicity, and that score rests on one assumption — that running two billing models is expensive. That assumption is the thing to check next.

Running it with AI​

AI is useful for two parts: proposing criteria you forgot, and drafting the justification line for each score from your documents. It should not choose the weights; those are your values, and outsourcing them is outsourcing the decision.

In MindMax, the option matrix is one of the built-in "converge" skills on the canvas. Give each option its own card, select all of them, and apply it: every option comes back scored against shared criteria in one table, each cell readable against its row and column. Then run an assumption audit on the row the matrix favours — most decisions turn out to rest on one premise nobody checked — and finish with a trade-off analysis and an action plan.

Where it goes wrong​

• False precision. A total of 29 versus 27 is not a decision. Close scores mean the matrix is telling you the choice depends on something it does not capture.
• Criteria chosen after the fact. If you added a column because your favourite option wins on it, you have built a justification, not a comparison.
• Unscorable factors. Some things — a founder's energy, a relationship — do not fit in a cell. Note them beside the table rather than forcing a number.

Part of thinking models on the canvas. The full workflow for a real choice is think through a decision. Previous in the series: six thinking hats. The series starts with the pre-mortem.

We said every quote in MindMax is checked against the source. This post is about why that sentence hides more than it shows — and what you should expect from any tool that claims the same.

A PDF is a drawing of text​

A PDF describes where to paint characters on a page. It does not, in general, record sentences or paragraphs. Reading text out of it is reconstruction, and the reconstruction has predictable seams:

• Hyphenation. "infor-" at the end of one line and "mation" at the start of the next. The sentence the model read says "information"; the raw text says "infor- mation".
• Ligatures. Typeset "fi", "fl" and "ffi" are often single glyphs. A word that looks like "efficient" can come out of the file as something that does not match "efficient" letter for letter.
• Page breaks. A paragraph that runs from page 12 to page 13 is two fragments in the file, with a page number, a running header and maybe a footnote in between.
• Columns and figures. Two-column layouts with a figure caption in the middle are where reading order goes wrong most often.
• Quotation marks and dashes. Curly versus straight quotes, hyphens versus en dashes. Models tidy these constantly.

Two ways to get it wrong​

A quote checker has two failure modes, and they pull in opposite directions.

Too strict, and real quotes fail. The model quoted the sentence correctly; the check rejected it because of a hyphen at a line end. Users learn that "not found" means nothing, and stop looking at the flag.

Too loose, and paraphrases pass. "The treatment may reduce risk" matches "the treatment reduces risk" closely enough to slip through. That is the failure the check existed to catch.

The job is to forgive the page — hyphens, ligatures, headers, quote styles — while staying strict about the words. A dropped "may" has to fail.

What we do about it​

Without going into the machinery, here is the behaviour you can see:

• Paragraphs are rebuilt across page breaks, so a quote is not cut in half by the bottom of a page or polluted by the running header above it.
• A quote is matched against the document's words, not its typesetting, so line-end hyphens and ligatures do not cause false alarms.
• A near miss is put back into the source's own wording. If the model tidied punctuation or swapped a word, the citation you see uses the document's exact text — and that exact text is what you are judging.
• A quote that cannot be found is never shown as checked. In an answer it is flagged at the end; an AI margin note whose quote does not match is sent back and redone.
• The citation lands on the sentence, highlighted on the page, so the last check — does this sentence support the claim? — takes one look.

What to expect from any tool​

If you are evaluating an AI reader, try this: take an answer, pick a quote that sits at a line break or a page break in the PDF, and click it. Does it land? Then ask about a passage you know is hedged, and see whether the hedge survives in the quote. Those two checks tell you more than any accuracy claim.

What is not claimed​

Extraction is not perfect. Dense mathematics, two-column layouts with interleaved figures and scanned pages with poor recognition are still where quotes are least reliable, and where clicking through to the page is worth the second. We are not claiming every quote is right; we are claiming a wrong one is shown as wrong.

The idea in full is checked quotes; why it matters is in why AI invents citations. The reader itself is on the AI PDF reader page.

Part of Thinking models in practice, a short series on classic methods and how to run them on your own material.

Where it comes from​

Edward de Bono introduced the method in his 1985 book Six Thinking Hats. He called the underlying idea "parallel thinking": instead of two people arguing from opposite positions, everyone looks in the same direction at the same time — all at the risks, then all at the value — so each angle gets a full hearing.

The six hats​

• White — what we know. Facts, figures, and what is missing. No interpretation.
• Red — what we feel. Gut reactions and unease, stated without justification. This hat legitimises instinct instead of letting it hide inside "objective" objections.
• Black — what could go wrong. Risks, weaknesses, reasons for caution.
• Yellow — what is the value. Benefits and why it might work, argued as seriously as the risks.
• Green — what else could we try. Alternatives, experiments, variations.
• Blue — what do we do next. Summarise, decide, plan the next step.

How to run it​

1. Frame one question. Specific enough that it could go either way.
2. Go through the hats one at a time. A few minutes each. Everyone wears the same hat at the same time.
3. Do not skip red. It is the one people find awkward, and often the one that explains why a meeting has been stuck.
4. End with blue. The output is a decision or a next step, not six lists.

The order is flexible; a common one is white, red, yellow, black, green, blue. Starting with black tends to kill ideas before yellow has had a turn.

A worked example​

Question: how can an open-source project with no marketing budget find its first 100 users?

• White: it solves one specific problem; there is no budget; the existing audience is unknown; "user" is not yet defined.
• Red: 100 feels achievable; direct outreach feels awkward; there is unease about being ignored.
• Black: broad messaging attracts the wrong users; launch attention fades fast; vanity metrics can hide poor adoption.
• Yellow: a narrow, useful tool spreads by referral; open development earns trust; a small cohort improves the product.
• Green: recruit around one painful workflow; turn the documentation into discovery; borrow distribution through integrations.
• Blue: define an active user, choose one segment, test for a short cycle, then widen.

The conclusion — fit before reach — came from black and yellow disagreeing in public instead of in someone's head.

Running it with AI​

The method suits AI well because the hats are fixed and the work is to fill them in. An AI will happily produce a red hat — "the first 100 feels achievable" — which a team might never say out loud. What it cannot do is know your feelings; treat its red hat as a prompt to say yours.

In MindMax, six thinking hats is one of the built-in "diverge" skills on the canvas, and it always comes back as six cards plus a summary that names the tension between them. Each line on a hat card has its own skill button, so you can dig into one risk, or run a red team on the most promising green idea, without disturbing the rest.

Where it goes wrong​

• It becomes a ritual. Six lists and no decision. The blue hat exists to prevent that.
• The question is too broad. "Our strategy" gives you six hats of generalities.
• Black dominates. In most teams caution sounds smarter than enthusiasm. Give yellow the same time.

Part of thinking models on the canvas. Previous: SCAMPER. Next: the decision matrix.

The usual approach — ask a chat assistant for one section at a time — produces sections that are each fine and a chapter that does not hold together.

Before anything else: the rules​

Universities and journals differ widely on what AI help is allowed in a thesis, and on what has to be disclosed. Some permit AI for editing but not for drafting; some require a statement of every tool used. Read your department's policy and talk to your supervisor first. Everything below assumes you are allowed to do what it describes, and that the ideas and the judgement are yours.

Why long documents drift​

A chat window has no memory of the chapter, only of the conversation, and conversations get long and get trimmed. So, section by section:

• Terms get redefined. "Institutional trust" is defined carefully in 2.1 and casually re-explained, slightly differently, in 2.4.
• Promises go unpaid. "We return to this in section 2.5" — and section 2.5 never does.
• The claim shifts. Each section is locally reasonable; together they argue something a little different from what the introduction said.
• Sources thin out. By the fourth section the model is writing from general knowledge, not from the papers you gave it.

None of these are sentence-level problems, which is why sentence-level polish does not fix them.

Step by step​

1. Write the one-sentence claim yourself​

Not the topic — the claim. "This chapter argues that X, because Y, against the view that Z." If you cannot write it yet, you are not ready to draft, and no tool will fix that.

2. Agree the outline before any prose exists​

An outline is where disagreeing is cheap. Each section should have the point it makes, not just a heading. Change it until each section earns its place. In MindMax, drafting does not start until you approve the outline, and a comment on one section replans that section rather than redrawing the whole outline.

3. Draft from your sources, with footnotes back to them​

Point the drafting at the papers and notes you actually read. Every claim drawn from a source should carry a footnote back to the passage, so your supervisor's "where does this come from?" has a one-click answer.

4. Keep a ledger of claims, terms and promises​

After each section, record what it argued, which terms it defined and what it promised to come back to. Before each section, check that list. You can keep it by hand in a note. In MindMax it is kept as the draft is written and handed to every later section, so a term is used rather than defined twice, and an unpaid promise is flagged in the final review.

5. Revise as tracked changes, never as a new copy​

When you ask for changes, they should come back marked on your text, grouped, with a reason — so you can keep the clarity fix and reject the sentence that softened your claim. (Why this matters: an AI should never rewrite your document silently.)

6. Hand it over as a real document​

A thesis chapter goes to a supervisor as Word or PDF, usually with equations and tables. Export from the same text you reviewed, so what they read is what you checked.

What AI should not do here​

• Decide the argument. The claim in step 1 is the thesis. Delegating it is delegating the degree.
• Be your fact-check. A consistency review checks that the draft agrees with itself and cites only what it was given. It does not tell you whether a source is right. Read the footnotes that carry weight.
• Supply sources you have not read. If a reference appears in the draft and you have not read it, it does not belong there.

Where it goes wrong​

• Voice gets close, not identical. Your supervisor knows how you write. Expect to rewrite the parts that matter most in your own words.
• Length is a budget, not a contract. A draft that runs short should deepen the argument, not pad.
• It is built for non-fiction. Argument and explanation, not narrative.

The mechanics are on book-length and long-form writing, the idea behind the ledger is the argument ledger, and the end-to-end workflow is thesis and long-form writing. Step by step: writing a long document.

Drafting is the easy part now. The expensive mistakes are elsewhere.

The mistake that costs you​

The worst outcome is not a clumsy answer. It is a fluent, confident answer that quotes last year's policy — the backup retention period you have since changed, the certification that lapsed, the encryption standard you migrated away from. An AI quotes a superseded document as happily as a current one; it has no way to know which one your company stands behind.

The second-worst outcome is an answer with no source at all, where the model filled a gap with what a typical company would say. It reads exactly like the others.

Both are invisible unless every answer has to show where it came from.

Step by step​

1. Clear out what should not be quoted. Move old policy versions and drafts out of the folder — or, in MindMax, right-click them and keep the AI out. Ten minutes here prevents the worst mistake.
2. Probe the three hardest questions first. Before drafting anything, ask the questions you are least sure of, such as "what do our documents say about data retention for backups?" A thin answer is a gap found while it is cheap.
3. Hand over the questionnaire as the structure. Every question kept in its own wording and order — not renamed, merged or reordered. Reviewers check answers against their own numbering.
4. Approve the outline before any answer is drafted. Every question and the point each answer will make. This is where you catch "we are going to say yes to SOC 2" before it is written down.
5. Check the answers that commit you. Click through the quotes for every certification, number, date and promise. Then look at the answers with no quote — those are the ones most likely filling a gap.
6. Mark it up and take only the edits you want. Comment where an answer is wrong, too confident or too long for the form. Revisions come back as tracked changes, grouped with reasons.
7. Export to Word. Each citation becomes a numbered source that names the document and keeps the quoted passage. Keep the Markdown copy: it is the starting point for the next similar questionnaire.

Why the quote is the whole point​

A questionnaire answer is a statement your company will be held to. "Where does it say that?" is a question you will be asked — by the reviewer, by legal, or by an auditor a year later. An answer with a quote from a named, current document answers it on the page. An answer without one is your company's signature on something nobody checked.

That is also why the line between "stands on our documents" and "does not" should be visible in the draft, not discovered later. An unquoted answer is not necessarily wrong. It is a task.

Where it goes wrong​

• Your documents are the ceiling. If the policy does not say it, the answer cannot honestly say it either. Gaps found here are often the most useful output of the whole exercise.
• Portal-based questionnaires still need answers pasted in by hand, one field at a time.
• Character limits. Long, well-sourced answers may need trimming for the form; do that as a revision you review, not by hand-deleting the source.
• Legal and compliance review still applies. This gets you a sourced draft fast. It does not sign off.

The full workflow, with the questions to ask at each step, is answer a questionnaire from your documents. How answers cite their sources is in chat with your documents, and how the citations survive into Word is on Word and PDF export.

Part of Thinking models in practice, a short series on classic methods and how to run them on your own material.

Where it comes from​

SCAMPER was arranged by Bob Eberle in the early 1970s from an earlier checklist of idea-spurring questions by Alex Osborn, the advertising executive who also popularised brainstorming. Its staying power comes from being a checklist: it does not ask you to be creative, it asks you seven specific questions.

When to use it​

SCAMPER is for an idea that exists and is not good enough yet — a product feature, a workshop format, a lesson plan, a process. It is the wrong tool for a blank page; for that, brainstorm first and then SCAMPER the version you like.

The seven questions​

• Substitute — what component, material, person or step could be swapped for another?
• Combine — what could be merged with something else, including another of your own ideas?
• Adapt — what has solved a similar problem elsewhere, and what would it look like here?
• Modify (or magnify/minify) — what could be bigger, smaller, more frequent, rarer?
• Put to another use — who else could use this, or for what else?
• Eliminate — what could be removed entirely without losing the point?
• Reverse (or rearrange) — what if the order, the roles or the direction were flipped?

A worked example​

Idea: a monthly one-hour internal talk where one team presents its work to the company. Attendance is falling.

• Substitute: replace the presenter with a demo by a user of that team's work.
• Combine: merge it with the onboarding programme, so new hires attend as part of their first month.
• Adapt: borrow the conference "lightning talk" format — five teams, ten minutes each.
• Modify: make it shorter and more frequent: twenty minutes, every two weeks.
• Put to another use: record it and use the recordings as documentation.
• Eliminate: drop the slides; show the actual product or the actual spreadsheet.
• Reverse: the audience brings questions in advance, and the team answers only those.

Not all seven are good. Three are worth trying, and two of them — lightning format and questions in advance — combine well. That is a normal SCAMPER outcome: seven prompts, a shortlist of two or three.

Running it with AI​

SCAMPER is close to ideal for AI: the questions are fixed, the work is generative, and you judge the results. The trap is generic output — "use AI to personalise it" under every heading. Give it the real idea with its real constraints, and ask for specific, testable variations.

In MindMax, SCAMPER is one of the built-in "diverge" skills on the canvas, and it keeps its shape: it always comes back as seven branches, one per question. Apply it to the version of the idea you like after a brainstorm, cross-breed the parts that survive, then distil a shortlist you picked yourself.

Where it goes wrong​

• Treating all seven as equally valid. Most will be weak. The value is in the one or two you would not have reached otherwise.
• Using it on a blank page. There is nothing to substitute or eliminate yet.
• Stopping at the list. An idea from SCAMPER is a hypothesis. Pick one and test the smallest version.

Part of thinking models on the canvas. Previous: first principles. Next: six thinking hats.

The usual voice-memo workflow fails at the second step. You record, the audio sits in an app you never open, and the transcript, when you finally get one, is a wall of "um, so, the thing is".

What makes capture work​

Three properties, in order of importance:

1. No context switch. If capturing means finding a window, the thought is gone. It has to be a shortcut that works over the browser, the spreadsheet or the call you are on.
2. You read it before it is saved. Transcription mishears names and jargon. A second to fix the one wrong word beats finding it a month later.
3. It lands somewhere you already look. A daily note in your own folder, as plain text — not a separate capture app you will have to export from.

Step by step in MindMax​

1. Press Option-Shift-Space (Alt+Shift+Space on Windows). A small capsule appears at the top of the screen, over whatever you are doing, and starts listening. Prefer to type? Just type.
2. Talk. Words appear as you speak. The language is detected from what you say, and punctuation is added as it goes.
3. Stop. Press Enter, or simply stop talking — after a pause there is a short countdown you can talk over. The transcript waits for you.
4. Fix the misheard word, then choose where it goes:
   • Enter saves it to the end of today's daily note, with the time, exactly as it reads. No AI involved.
   • ⌘Enter (Ctrl+Enter on Windows) hands it to the AI to file into the note you have open.
5. Esc hides the capsule and keeps the draft for next time. Only the × throws it away.

What "AI filing" does — and does not do​

When the AI files a quick note, it changes how you said it, not what you said:

• Typos, misheard words, grammar and word order are fixed, even in a one-line note. Filler, false starts and repetition go. A correction keeps only the corrected version. The point comes first. Something you have to do becomes a task.
• Every point, name, number and hedge stays. Nothing is added. A question stays a question — it does not answer it for you. A correction it is not sure of is listed for you to check.
• It finds where the thought belongs within that note: under its heading, after the item it continues, as a list item if that section is a list. It does not go looking for a different note.
• It comes back as a marked change you accept. Nothing is written until you do.

You can also reorganise a note by voice: "rename this heading to Launch week", "merge the two printer items", "move that line under Today". Ask it to "turn the above into an email" and it is kept as a task with an offer to run it — new writing only happens when you say so, as its own change.

Where it goes wrong​

• Names, jargon and noisy rooms are where transcription slips. That is why it always stops for you to read.
• It is for thoughts, not meetings. No speaker separation, no hour-long transcripts; a single recording stops at ten minutes.
• Desktop only. The thought you have on a walk still goes into your phone. Paste it into the capsule when you are back.
• With your own API key, transcription needs a speech model at your provider; if it has none, pick one for voice in model settings.

Everything about voice notes and AI quick notes, with the step-by-step in voice and quick notes. Where a week of quick notes ends up: notes that write back.

Part of Thinking models in practice, a short series on classic methods and how to run them on your own material.

Analogy versus first principles​

Most reasoning is by analogy, and that is fine: "we did it this way last time", "this is how others in the industry price it". Analogy is fast and usually right.

It fails when the conditions that made the old answer right have changed — and nobody has noticed, because the answer is inherited rather than derived. First principles is the method for checking whether an inherited answer still follows from anything.

The idea is old: Aristotle described first principles as the propositions from which a thing is known and which are not themselves derived from anything else. The modern, practical version is less grand: which parts of this are actually required, and which are just habit?

How to do it​

1. Write down the current answer and the reasons usually given for it.
2. Label each reason. Is it a constraint (a law, a physical limit, a hard cost), a requirement (something the user genuinely needs), or a convention (how it is done)?
3. Set the conventions aside. Not delete — set aside.
4. Rebuild from constraints and requirements only. What is the simplest thing that satisfies them?
5. Compare with the current answer. Where they differ, either the convention carries a reason you missed (put it back, labelled properly) or it is the opportunity.

A worked example​

Problem: a small research group spends two weeks every year preparing a reading list for new students.

Usual reasons: every student needs a canonical list; the list must cover the whole field; a senior member must curate it.

Labelled: "must cover the whole field" is a convention — students need to start working, not survey the field. "Senior member must curate" is a requirement only for the first five papers, which set the frame. The real constraint is the students' time in their first month.

Rebuilt: five foundational papers chosen by a senior member, plus a standing folder of the group's own annotated reading that students extend themselves. Two days instead of two weeks, and the list keeps itself current.

Running it with AI​

AI is useful at step 2, because it is good at noticing that a "must" is really a habit, and at step 4, because it can generate several rebuilt answers quickly. It is weaker at knowing which of your constraints are real — it does not know your budget, your regulator or your users. Give it those explicitly.

In MindMax, first principles is one of the built-in "dig deeper" skills on the canvas. Apply it to a card holding the problem, ideally with your notes or documents attached, and it strips away the analogy and convention and reasons up from what is left, one branch per step. Aim it at a single line of a card to work on one assumption without disturbing the rest.

Where it goes wrong​

• Everything is labelled a convention. Some conventions encode hard-won lessons. If you cannot say why a convention exists, find out before discarding it.
• It is used where analogy would do. For most decisions, "what worked last time" is good enough. Save this for the ones where it might not be.
• The rebuilt answer is never tested. A derivation is a hypothesis. Try the smallest version of it first.

Part of thinking models on the canvas. Previous: the steelman. Next: SCAMPER.

The third time you type the same instruction into a chat box — "decisions first, each with an owner and a date" — that sentence has become a skill. It just has not been saved yet.

Skills versus prompts versus custom instructions​

• A saved prompt is text you paste. It works if you remember it exists.
• Custom instructions apply to every message. Good for standing preferences ("British spelling", "be brief"), bad for one kind of job, because they apply to every other kind too.
• A skill applies to one kind of job, is called by name or picked when the job fits, and lives as a file you can read, edit and send to someone.

The format has spread across AI agent tools because it is simple: a Markdown file, a short header, and prose.

The anatomy​

---
name: meeting-notes
description: Use when turning rough meeting notes or a transcript into a record of decisions and actions.
---

1. Lead with decisions. Each one gets an owner and a date. If no date was agreed, write "no date agreed"
   rather than inventing one.
2. Then open questions, each with who is expected to answer.
3. Then a short narrative only if something cannot be expressed as a decision or question.
4. Never attribute a statement to a person unless the notes name them.

Two parts, and they do different jobs.

The description decides whether the skill is picked. When an AI chooses a skill automatically, it usually sees only names and descriptions until one is needed. So write the description as the situation it applies to — "use when turning rough notes into decisions" — not a summary of what is inside it.

The body decides what happens once it is picked. Write it as instructions to a careful colleague who will follow them literally. Concrete beats clever: "if no date was agreed, write no date agreed" does more than "be accurate about dates".

Five rules that make skills work​

1. One job per skill. "Meeting notes" is a skill. "Writing helper" is a wish.
2. Say what to do when information is missing. That is where models improvise, and improvisation is where they invent things.
3. Describe the output's shape. A list, a table with these columns, four quadrants. Shape is easier to follow than tone.
4. Keep the shelf small. Overlapping descriptions make the right skill harder to pick. Retire the ones you never use.
5. Fix the line, not the whole file. When a result is wrong, the instruction usually did not say what you meant. Change that one line and run it again.

Read before you install​

A skill contains no code, but text can still tell an AI to do things you would not choose. Read a skill from someone else in full before installing it, the way you would read a script before running it.

In MindMax​

Skills are files in your folder: drop a SKILL.md into the skills folder and it counts; move it out and it stops. Type / to call one by name, or describe the task and let the matching one be picked. Typing /new-skill and describing the method in plain words gets you a drafted skill to read and approve — nothing is saved until you say so.

The same skill can be switched on for chat, for the canvas and for the editor. In chat it can take several turns and read your documents. On a canvas card, its output comes back in the shape the method calls for, such as a table or four quadrants. In the editor it runs on the paragraph you selected and comes back as a tracked change.

SKILL.md files written for other agent tools work as chat skills, with one boundary: only the text carries over. Scripts or other files in a skill's folder are not run, and steps that depend on tools MindMax does not have will not happen.

What a skill cannot do​

A skill is instructions, not new abilities. It changes how a job is approached using what the AI can already do — read your documents, search, write files. If you were hoping to add a capability, this is the wrong mechanism.

More on AI skills, with the step-by-step in using skills. The definition on its own is in the glossary: agent skills. The built-in ones are mostly classic methods — see thinking models.

Slides are a strange thing to revise from. They are the lecturer's prompts, not the lecture. The half that explained them lived in someone's head and is gone.

Step 1 — Rebuild the missing half, slide by slide​

Summarising a deck is the least useful thing an AI can do with it: the bullets already are a summary. What you need is what was said around them — the explanation, how this slide follows the last one, and what you should be able to answer about it.

In MindMax you open the .pptx and start a reading on it. Each slide gets a card: what the speaker would have said, how it connects to the slide before, a question worth asking, and whether you can move on quickly. Speaker notes, where they exist, are part of what it reads — on decks people actually gave, the notes are often where the argument is.

Step 2 — Box the part you did not follow​

On a slide, the thing you are stuck on is rarely a sentence. It is the arrow from one box to another, the third term in the equation, the quadrant of a chart. Describing it in words is harder than the question.

Drag a box around it and ask. The words and numbers under the box come with the question from the file itself, so the answer reads values from the slide instead of guessing from a picture. The rest of the deck is in reach, so "how does this connect to the definition earlier?" has something to answer from.

Step 3 — Put the course in one folder​

The deck alone is thin. Put the lecture notes, the textbook chapter and last week's slides in the same folder, and a question about one of them can be answered from the others. Every answer that uses a slide cites it by number; one click opens the page with the passage highlighted.

Step 4 — Stop reading, start answering​

This is the step that matters and the one most people skip. Re-reading explanations produces a feeling of familiarity that is easy to mistake for knowing. Answering questions without looking — retrieval practice — is one of the best-replicated findings in learning research, and it is uncomfortable for exactly the reason it works.

A few ways to do it with AI:

• Self-quiz. Ask for questions on a lecture, answer them, and only then see the marking. In MindMax's canvas, the self-quiz skill marks the ones with definite answers on the spot and reports by what you got wrong — and flags the ones you were confident about, which are the ones worth knowing.
• Explain it back. Write the idea in plain words and have the gaps pointed out (the Feynman technique).
• Check what it assumes. Rate yourself on the prerequisites a topic leans on, so the gap is found before it trips you in the exam.

Step 5 — Space it out​

Come back to the questions you got wrong two days later, then a week later. The cards and quizzes are saved in your folder, so the second visit costs nothing and starts where you stopped.

Where it goes wrong​

• The AI was not in the room. It reads the slides and notes, not the lecture. If the lecturer made a point that is on no slide and in no note, it cannot know. Check against your own notes of what was said.
• A slide with three words gives it little to work with. The honest card is a short one. Adding the textbook chapter to the folder helps more than asking again.
• Old .ppt files do not open; save them as .pptx first. A deck exported as images needs recognising before anything can read it.
• Understanding is not memorising. Formulas you need to reproduce under time pressure still need practice by hand.

Everything the slide reader does is on lecture slides with AI; the learning skills are on thinking models. The end-to-end version is study for an exam from slides. Download MindMax and open this week's deck.

Part of Thinking models in practice, a short series on classic methods and how to run them on your own material.

Why it is worth the effort​

Beating a weak version of the other side feels like winning and teaches you nothing. The people you need to persuade — reviewers, colleagues, the reader of your essay — already know the strong version. If your argument only answers the weak one, they will notice before you do.

The idea has a well-known formulation in what the philosopher Daniel Dennett called Rapoport's rules: before criticising a view, restate it so clearly that its holder would say "thanks, I wish I had put it that way", list where you agree, and mention anything you learned from it. Only then rebut.

How to do it​

1. Name the view precisely. Not "the critics" — the specific claim, ideally in a sentence its supporters would sign.
2. Find its best evidence. What would a thoughtful supporter point to? Their strongest data, not their loudest slogan.
3. Find its best reasoning. Which values or assumptions make it compelling to intelligent people? A view held by smart people usually rests on something real.
4. Write it out as its advocate would. In the first person, without sarcasm, a paragraph or two.
5. Check the test. Would someone who holds this view recognise it and be satisfied? If not, revise.
6. Only now, respond — to that version.

A worked example​

Your position: remote-first teams should drop most recurring meetings in favour of written updates.

The steelman of the other side: "Written updates are excellent for reporting what happened and terrible at surfacing what nobody thought to write down. The value of a recurring meeting is not the agenda; it is the two minutes at the end when someone says 'one more thing' — the half-formed worry that would never survive being written. Removing meetings removes the cheapest channel for weak signals, and the cost does not show up until a problem that was mentioned in passing in March becomes a crisis in June."

That is a better objection than "people like meetings", and it changes the proposal: keep written updates, but protect one short, agenda-free slot for the things nobody would write down.

Running it with AI​

Steelmanning is one of the best uses of an AI, because it has none of your attachment to your position and it has read a great deal of the other side. Two conditions:

• Point it at the actual argument. Give it the paper you are disagreeing with, not a description of it. A steelman built from the source text is grounded; one built from a summary drifts toward the generic.
• Ask for the steelman before the critique. If you ask for both at once, you mostly get critique.

In MindMax, steelman is one of the built-in critique skills on the canvas. Apply it to a card holding a claim — or to a card pointing at the paper you are arguing with — and it builds the other side at its strongest before anything else happens, citing the passages it drew on. Then run an assumption audit on your own position and see which one survives.

Where it goes wrong​

• The steelman becomes a different argument. Strengthening a view is not replacing it with one you find easier to agree with. Check it against what its advocates actually say.
• It turns into false balance. Steelmanning a view does not make it right. Some positions are weak at their strongest, and saying so after a fair hearing is the point.
• You never come back to your own view. The exercise is a step toward a position, not a substitute for having one.

Part of thinking models on the canvas. Previous in the series: the pre-mortem. Next: first principles.

On a paragraph, a rewritten copy is a small annoyance. On a chapter, it is the reason people stop letting AI touch anything that matters.

The problem with "here is the improved version"​

Ask most assistants to tighten a section and you get a fresh copy of it. It reads well. It is also a different text, and somewhere in it:

• a hedge you chose deliberately ("may", "in most cases") is gone,
• a citation or footnote quietly dropped out,
• a sentence that was fine has been rephrased for no reason,
• a term you defined carefully has been swapped for a synonym.

To find those you have to compare the two versions line by line — longer than making the edits yourself. So people do one of two things. They accept the whole copy and hope, or they stop asking.

Neither is a good outcome for a tool whose job is to save you time on writing that matters.

What "edits wait for you" means​

The rule we built around is simple to state: an AI may add to your files, but may not replace what you wrote on its own.

• A change to existing text — a rewrite at the cursor, an edit you asked for in a conversation, a revision from comments you left — comes back marked in place and is written only if you accept it.
• New material — a note the assistant creates, a translation saved beside the original — is written directly, because asking permission for every new file is friction with no safety in it.
• Every write the AI makes is kept as a version you can restore.

The first point is the one that changes how editing feels. If unmarked text is certain to be untouched, you only have to read the marks.

What makes tracked AI edits actually usable​

Marking the changes is necessary but not enough. A revision of a whole chapter can easily be a hundred small edits, and a hundred individual accept buttons is its own kind of unreadable. What helped:

Group the changes and say why. A revision arrives as groups — "tighten the opening", "make the terminology consistent", "answer comment 3" — each with a one-line reason. You judge the intention once, then skim the marks that carry it out.

Flag what was lost. Footnotes, citations and links are meant to survive an edit. A group that drops one anyway carries a warning, so you see the loss before accepting it, not after sending the document.

Do not write until you say so. Your file is untouched while you review. Only the groups you accept are written.

Keep applied changes reversible. Accepted is not final. A change you regret tomorrow is one step back in the document's history.

Where it does not apply​

• Your own typing is not tracked. This is about how AI edits reach you. What you type goes straight into the file, and into its history.
• A cursor suggestion is one change. Select a whole chapter and ask for a rewrite at the cursor, and you get one change too big to judge. For that, describe the change in a conversation or leave comments, and it comes back as many small groups.
• Export is not Word's revision mode. Exporting a Markdown document to Word gives you the accepted text, not a file full of revision marks. A Word file you ask the AI to revise is different: the same review happens in MindMax, and the edits you accept are written into the .docx as Word tracked changes — see how to have AI edit a Word document with tracked changes.
• Judgement is still yours. Marks tell you what changed, not whether the change is right.

How it looks in practice is on tracked AI edits; the rule itself is written up as edits wait for you, and the walkthrough is reviewing AI revisions. If you have a draft and a stack of feedback, revise a draft from feedback is the workflow.

The catch comes first, because it decides whether the rest applies to you: books bought from the Kindle store carry DRM, and they cannot be opened. That lock is not ours to remove, and this post does not explain how to remove it.

Which Kindle files will open​

• Will open: DRM-free .mobi and .azw3 files — documents you converted yourself, books from publishers and authors who sell without DRM, public-domain collections, files a colleague generated.
• Will not open: books downloaded from the Kindle store to the Kindle app or a device. They are encrypted to your account.
• Opens, with limits: older .mobi files that never had a table of contents. The text reads fine; there is nothing to navigate by chapter, which is where most of the value comes from.

If you have the choice when buying, an EPUB without DRM is the most portable thing you can own.

Step by step​

1. Put the file in a folder. Anywhere you already keep reading material. There is no library to import into.
2. Open it. The first time, the Kindle file is converted into a proper EPUB behind the scenes. Your original .mobi or .azw3 stays exactly as it was.
3. Check the chapters. The contents panel comes from the table of contents the book was published with, not from guessing which lines look like headings. If the chapters look right, everything else will work.
4. Read normally. Reflowable text at your font size, highlights that stay anchored to the passage, and search inside the book that lands on the passage, not just the chapter heading.
5. Ask about it. Select a passage and ask, or ask about the whole book. Answers quote the passage they rest on; click a quote and the book jumps there with the sentence highlighted.
6. Or let it read with you. Pick a lens — guide, review, critique or study — and the AI works through the book a chapter at a time, leaving notes in the margin beside the passages they are about, and a closing write-up when you stop.

Why chapter structure matters so much​

Most books are longer than any model can hold at once, so something has to decide what gets read. A tool that cuts the book into fragments and searches them will answer from whichever pieces matched your wording — and miss the argument that runs across chapters.

A book with a real contents page can be navigated the way you would: look at the chapters, read the ones the question needs in full, and go back for another if the first did not settle it. That is why step 3 is worth ten seconds of your attention.

Where it goes wrong​

• DRM. Covered above. If it is from the Kindle store, it will not open.
• Badly built books. Broken chapter metadata, or a whole book in one giant file, leaves chapter navigation with nothing to work with.
• Fiction. The reading lenses are built around argument and evidence. It will summarise a novel; it has nothing useful to say about whether the novel is any good.
• A whole-book pass is a real job. Opening, reading and searching your books costs nothing; a pass over four hundred pages is AI work and is priced like it. Try one book before a shelf.

Everything the reader does is on the AI EPUB and Kindle reader; what each lens marks is on annotation lenses. Why the AI leaves fewer notes than you might expect is in the AI annotates your book sparsely, on purpose. Download MindMax and open a book you have been meaning to finish.

That split is worth knowing before you start, rather than four hours into a job that will not finish.

Why people want this​

The reason is usually one sentence: nothing may leave the machine. A contract under NDA, patient notes, an unpublished manuscript, a client's financials. For that material, "the text goes to a provider with a good privacy policy" is not the same answer, and it is right to treat it differently.

A model running locally is the only configuration where the honest answer is "nothing is sent anywhere". Bringing your own cloud API key removes one party from the path; it does not remove the path.

What a local model is good at​

Laptop-sized models have improved fast, and for a lot of everyday document work they are genuinely fine:

• Explaining a selected passage — the text is short, the question is local.
• Questions about one open document that fits comfortably in the model's context.
• Tidying, rephrasing and summarising a note you wrote yourself.
• Quick edits where you will read every change anyway.

For these, speed is the main cost, and on a recent machine it is usually acceptable.

Where it struggles​

Three properties separate the big jobs from the small ones, and laptop-sized models are weakest at all three:

1. Long context. Reading a book well means holding a lot of text at once. Local models often run with a much smaller context than their cloud counterparts, or slow down sharply as it fills.
2. Dependable tool use. Navigating a book — open the contents, pick the chapters, read them, go back for another — is a sequence of tool calls. A model that gets one call slightly wrong derails the whole sequence.
3. Staying consistent over hours. A three-hundred-page translation or a long draft is hundreds of calls that all have to follow the same instructions. Small models drift.

So: an AI reading pass over a whole book, whole-book translation and book-length drafting are the jobs we would not start on a local model. It will be slow at them, or it will not finish.

Reading images is a separate question​

Asking about a boxed chart, recognising a scanned page or reading a photographed whiteboard needs a model that takes pictures. Many popular local models are text-only. If yours is, those features have nothing to run on until you point that job at a model that reads images.

Splitting the work​

The practical setup for most people is not all-or-nothing. MindMax groups model work into tiers — quick work, everyday work, the passes where being right matters, and reading images — and you choose which model sits in each, with per-task overrides.

• If nothing may leave the machine, point every tier at the local model and accept that the long jobs are out of reach. Keep those documents in their own folder so you always know which rules apply.
• If some material is sensitive and some is not, keep the local model for the sensitive folder's everyday questions and use a cloud model for long jobs on everything else.
• If you just want to save money, a local model for quick work is a reasonable place to start. Measure before assuming it is cheaper overall; your time is part of the bill.

Local providers need no key. Requests go to an endpoint on your own machine, and nothing about them reaches us.

What is not claimed​

We have not published benchmarks for specific local models, and capability changes month to month. The claim here is about the shape of the work: short and contained is fine, long and multi-step is where a laptop-sized model runs out. Try the job you actually need on a short document first.

Setting it up is in use your own API key, and the full picture — including where your own key is and is not the same as privacy — is on bring your own key. The reasoning behind making every feature work on your own key is in the subscription is a convenience, not a gate.

This is the first post in a short series, Thinking models in practice: one classic method per post, what it is for, how to run it by hand, and how to run it on your own material.

Where it comes from​

The psychologist Gary Klein popularised the pre-mortem in a short Harvard Business Review piece in 2007. It rests on an older finding about "prospective hindsight": imagining that an event has happened makes it easier to generate explanations for it than imagining that it might happen. "Why might this fail?" invites reassurance. "It failed — why?" invites a story.

When to use it​

• A plan is about to be approved and everyone is quietly enthusiastic.
• You are the person who proposed it, and you know you are not the right one to find its flaws.
• The failure would be expensive and slow to notice: a launch, a hire, a migration, a move.

It is less useful early, when the plan is still vague. You cannot imagine the failure of something that has no shape yet.

How to run it​

1. State the plan in one sentence, with a date. "We launch the paid tier to existing users on 1 March."
2. Declare the failure. "It is a year later. The paid tier was a disaster." No hedging.
3. Everyone writes reasons silently, for a few minutes, before anyone speaks. This step is the method; discussion first means the most senior voice sets the list.
4. Go round, one reason each, until the reasons run out. Write them all down. No rebuttals yet.
5. Pick the two or three that are both plausible and serious, and decide what would change in the plan because of them — a check, a trigger, a smaller first step.

Alone, you can run steps 2–5 on paper. The silent-writing step matters even more: your first reason will be the one you were already worried about, and the useful ones come after it.

A worked example​

Plan: an open-source project with no marketing budget aims to reach its first 100 active users in three months by posting on two developer forums.

It failed. Why?

• The posts landed, but the install took forty minutes on a clean machine, and most people gave up.
• "Users" was never defined, so the team celebrated 400 stars while active use stayed in single digits.
• The one forum that worked attracted people with a different problem; their feature requests pulled the roadmap sideways.
• The maintainer burned out answering support questions in week three.

The changes: define an active user before launch, record a clean install from scratch, and pick one narrow audience instead of two broad ones.

Running it with AI​

AI is good at the part people find hardest — generating many specific failure stories quickly — and bad at knowing which ones matter for you. Two rules keep it useful:

• Give it the real plan and the real context, not a topic. "A launch" gets generic risks. Your plan document, your numbers and your constraints get specific ones.
• Keep the selection step human. The AI produces candidates; deciding which two to act on is yours.

In MindMax, the pre-mortem is in the canvas's starter library — install it once, then apply it to a card holding your plan, with the planning documents attached. Each failure story comes back as its own branch, which you can expand, drop, or turn into a check. The built-in red team is a close cousin: instead of "it failed, why?", it attacks the plan from an adversary's side.

Where it goes wrong​

• It becomes a list of everything. Twenty risks with no selection is anxiety, not planning. End with two or three changes.
• The plan is too vague. "Grow the business" fails for every reason at once.
• Nobody acts on it. A pre-mortem that changes nothing in the plan was a meeting.

More methods like this: thinking models on the canvas, and the definition in the glossary: mental model. Next in the series: the steelman.