An Author's Guide to AI-Assisted Writing

Run this before you read anything, on a topic you are about to spend weeks in. You are looking for the shape of the conversation, not for facts you will cite.

I am researching [topic] for a nonfiction book aimed at [reader].

Give me the landscape, not a summary. Specifically:
1. The four or five positions people actually hold on this, stated as their strongest advocates would state them.
2. For each position, the single work most associated with it, with author and year.
3. What the mainstream view was ten years ago and how it changed.
4. Two questions where the evidence is genuinely unsettled.
5. What a well read person in this field would consider a beginner's mistake.

Mark anything you are unsure of with [UNCERTAIN]. Do not fill gaps with plausible titles. If you do not know a specific work for a position, say so.

The single most useful research artifact in a nonfiction project. Do it once per chapter, at the end, before you think you are finished.

Here is a chapter draft. Extract every factual claim into a table.

Columns: the claim in my words, the type (statistic, date, quotation, attribution, general fact), whether I gave a source in the text, and a verification difficulty rating of easy, moderate or hard.

Include claims I have stated as obvious. Those are the ones that get me in trouble.
Do not verify anything. Do not tell me whether the claims are true. Just extract them.
Do not rewrite or edit my text in any way.

This is the one people skip. You wrote it, so you believe it. Ask the machine to try to break it instead of trying to help it.

Read this chapter as a hostile fact checker at a publishing house whose job is to find the thing that gets us sued or embarrassed.

List, in order of risk:
1. Claims stated with more confidence than the underlying evidence supports.
2. Numbers with no source, including ones that sound right.
3. Attributions of a quote or idea to a person that may be misattributed.
4. Anything that was true when commonly reported but may have changed since.
5. Any place I have implied causation from what is probably correlation.

For each, tell me what specifically you would want to see before you let it print.
Do not correct the text. Do not suggest replacement wording.

Before you defend a position in print, find out who has already attacked it and how well.

My book argues [state your claim in one sentence].

Who has argued against this, and what is their strongest version of the argument? Give me the three most serious objections, each stated at full strength with no hedging and no straw men, plus who is associated with each. Then tell me which of the three I would have the hardest time answering, and why.

For terms of service, platform rules, submission guidelines, contest requirements. Paste the page text rather than the link when you can.

Below is the current text of [platform]'s policy on [topic], pasted from their site today.

Tell me:
1. What it explicitly requires, quoted.
2. What it explicitly forbids, quoted.
3. What it does not address at all, which is the part I care most about.
4. Any place where the wording is ambiguous enough that two reasonable people would read it differently.

Answer only from the text I pasted. Do not use anything you remember about this policy from training, because policies change and your memory of it is dated. If the text does not answer a question, say the text does not answer it.

For the moment when two credible accounts do not match and you have to write one sentence.

Here are three sources on [event or question]. They do not agree.

Build a table: the point of disagreement down the left, what each source says across the top.
Then, separately: for each disagreement, tell me what kind of disagreement it is. Different data, different definitions, different time periods, different methodology, or genuine dispute.

Do not tell me which source is correct. Tell me what a reader would need to know to judge for themselves.

Useful in nonfiction for historical chapters, and useful in fiction when you are working from real events.

From the material below, extract every dated event into a chronological table.

Columns: date as given, event in under fifteen words, source, and confidence.
Confidence is one of: stated explicitly in the source, inferred by me from context, or unclear.
Where two sources give different dates for the same event, give it one row and note both.
Where a date is given only as a month or year, keep it that vague. Do not resolve it to a day.
Add nothing that is not in the material.

The main event. Budget an hour. Do not answer in the chat window if you can help it.

I am writing a book about [subject]. I have been thinking about it for [time] and I do not yet know what it is.

Your job is to interview me until I find out. You are not writing anything. You are not summarizing. You are not offering ideas.

Rules:
Ask one question at a time and wait for my answer.
Follow up on my answer before moving on. If I am vague, say I was vague and ask again.
If I give you an abstraction, ask for the specific case it came from.
If I give you a specific case, ask what it generalizes to.
Push on anything I say that sounds like something I read rather than something I learned.
Do not compliment my answers.
Do not tell me my idea is interesting.
Do not suggest what I might mean. Ask.

Start with what I would say to a person at a dinner party who asked what the book is about, and go from there. Keep going until I ask you to stop.

Ask it early, when it is cheap. It is much more expensive at thirty thousand words.

Here is my book idea in a paragraph: [paste].

Make the strongest possible case that this book should not be written. Consider: it already exists, the audience does not want it, the premise is wrong, it is an article and not a book, and I am not the person to write it. Take each seriously and give me the sharpest version.

Then, in one sentence, tell me what would have to be true for the book to be worth writing anyway.

Do not soften it and do not end on encouragement.

Different from the counterargument prompt under Research and verification. That one finds published opposition. This one finds the person in the room.

Play a reader who bought my book, is 40 pages in, and is starting to disagree with me. They are intelligent, they are not hostile, and they have relevant experience of their own.

Here is my argument so far: [paste your summary].

Ask me the questions that reader would be asking. One at a time. Do not resolve them for me. Keep pressing on the ones where my answer is thin, and tell me when my answer is thin.

The prompt that finds the book you can write, as opposed to the book you would like to have written.

Interview me about what I know from experience rather than from reading, on the subject of [topic].

For every claim I make, ask me how I know it. Ask for the specific instance. Ask what year, what happened, what I expected and what actually occurred.

Flag anything I say that I could only have gotten from a book, an article, or general knowledge, and set it aside. I am trying to find what is only mine.

At the end, list the things I told you that you would not have been able to write yourself.

If you cannot answer this you do not have a book yet, and finding that out in an hour is a gift.

Ask me these five questions one at a time, and push back if my answer is vague or if I am describing a topic instead of a change.

1. What does the reader believe on page one?
2. What do they believe on the last page?
3. What can they do afterward that they could not do before?
4. What is the cost to them of not reading it?
5. Which of those four answers is the actual promise, the one that goes on the cover?

Do not accept a topic as an answer to any of them.

Vagueness here shows up later as a description that sells nothing.

I will describe my intended reader and you will keep narrowing me until the description is a person rather than a demographic.

After each of my answers, ask one question that makes the picture more specific. Age, situation, what they have already tried, what they have already read, what they are embarrassed about, what they would say to a friend about why they have not solved this yet.

Stop when you can tell me, in three sentences, what this person types into a search box at eleven at night. Then tell me those three sentences and ask if I recognize them.

The plain version. Use it on any outline, any genre, at any stage.

Here is the outline for my [genre] book. [paste]

Do not rewrite it. Diagnose it.

1. Where does the argument or story lose momentum, by chapter number?
2. Which chapter is doing the least work?
3. Where am I asking the reader to accept something I have not yet earned?
4. What is promised early and never delivered?
5. If I had to cut fifteen percent, what goes, and what breaks when it does?

Answer as a numbered list. No preamble, no summary of my outline back to me, no encouragement.

Models are good at this because it is a comparison task, not a creative one.

Here is my chapter list in current order: [paste titles and one line each].

Propose two alternative orderings. For each, give me the single sentence reason it is better, and the specific cost of the change, meaning what a reader loses by not getting the original sequence.

Then argue for keeping my original order as strongly as you argued against it. I want the argument on both sides, not a recommendation.

The most common structural fault in a first outline and almost impossible to see from inside.

Here is my outline. For each chapter, state in under ten words the single job it does for the reader.

Then group the chapters by job. Show me every group with more than one member.

For each such group, tell me whether these are genuinely different angles on one idea, in which case they should be one chapter with sections, or one idea I have not yet noticed I am repeating.

Do not propose new chapters. Do not write any chapter content.

Chapter 7 covers why this differs from the outline you write for yourself. This prompt produces the machine readable version, which is what you paste into every later diagnostic session so the model knows where it is.

Convert my outline into a structured reference document for use as context in later sessions.

For each chapter: number, title, the one sentence job, the reader's state before and after, the three to five items it must contain, what it depends on from earlier chapters, and what later chapters depend on from it.

Use consistent headings so I can paste any single chapter block on its own.
Keep my language for the ideas. Do not invent content I did not give you. Where my outline is silent on one of these fields, write [NOT SPECIFIED] rather than filling it in.

The one that finds the hole. Expect at least one answer that annoys you.

Below is my outline and the promise of the book in one sentence.

Assume the promise is correct and the outline is incomplete. What is missing?

Specifically: what question does this outline raise in a reader's mind and never answer, what step does it skip between two chapters that are further apart than they look, and what objection does it never address.

Give me at most three candidate missing chapters, each with the reader question it answers and where it would sit. If you think nothing is missing, say so and defend that.

The brief is what you read before you talk a chapter. This turns an outline into a stack of them.

For each chapter in this outline, produce a chapter brief in exactly this format:

Reader believes at start:
Reader believes at end:
The one thing this chapter does:
Items it must contain:
What it must not do, because another chapter does it:

Draw only on what is in my outline and notes. Where you cannot fill a field from what I gave you, write [MISSING] and move on. Do not write chapter content, sample paragraphs, or opening lines.

Do this once, before drafting, and again after a big structural change. It catches the forward reference problem that no amount of rereading catches.

Build a dependency map from this outline.

For each chapter, list the concepts, terms, characters or events that a reader must already have to understand it, and where each is introduced.

Then flag every case where a chapter uses something introduced later, and every case where something is introduced and not used again for more than three chapters.

Return it as a table. Do not suggest fixes. I will decide whether to move the introduction or the use.

The fix for the counting problem is structural. Give the model something it can see itself completing.

I have a chapter that needs to be roughly 4,000 words. You cannot count words and I am not going to ask you to.

Instead, from my chapter brief below, propose a list of eight to twelve beats that would fill that space: each one a specific thing that happens or gets established, in one line, in order.

Do not write the beats out as prose. Do not estimate word counts per beat. I want the list only.

Run this before you try to add twenty thousand words, because adding words to existing scenes cannot fix a book that is missing a complication.

My manuscript is [current] words and should be around [target]. Here is the outline and a chapter list with word counts.

Tell me which of these two problems I have, and give evidence from the structure rather than from the prose:
Under-plotted, meaning too few events, complications or arguments for the length I want.
Underwritten, meaning the right events are present but rendered too thinly.

If it is under-plotted, tell me where the structure has room for a complication that is not there.
Do not tell me to expand anything yet.

"Expand this" gets you adjectives and weather. Name the mechanism instead.

Here is a scene or section that is too thin. Do not expand it.

Tell me which of these five mechanisms it is actually missing, and where exactly, by quoting the sentence each opportunity sits next to:
1. Interiority, meaning what the point of view character thinks and does not say.
2. Side character presence, meaning someone else in the room having their own reaction.
3. Sensory detail that carries emotional state rather than decorating the room.
4. Action beats replacing dialogue tags.
5. A reaction scene, meaning the aftermath that is currently skipped.

Rank them by how much the passage would gain. Then stop. I will write the additions.

Budget one to two hours of human editing per thousand words of assisted output. This prompt makes you look at that number.

Build me a schedule for a [word count] [genre] book with these constraints: I have [hours] a week, my deadline is [date], and I have [amount] drafted.

Use these assumptions and show them in the output:
A forty minute talking session produces roughly 4,000 to 6,000 raw words, of which 3,000 to 4,000 survive cleanup.
Revision requires 1 to 2 hours of human editing per 1,000 words of draft, and this number does not shrink.
Add a 20 percent contingency on the revision figure, not the drafting figure.

Return a week by week table. If the deadline is not achievable under these assumptions, say so in the first line and tell me what would have to change: scope, date, or hours.

Cutting is a classification job, which models do well. Deciding is yours.

Here is my manuscript outline with word counts per chapter, and the promise of the book in one sentence.

I need to cut [number] words. Give me a ranked cut list: what to cut, how many words it is worth approximately, what the book loses, and what the book gains.

Rank by ratio of gain to loss, not by size. Include at least two items you expect me to refuse.
Do not perform any cuts. Do not rewrite anything.

The flagship prompt of this book. Paste your Voiceprint above it, always. This is the one that turns forty minutes of talking into a chapter, and every constraint in it exists because leaving it out produced something worse.

This is a transcript of me talking through a chapter of my book. Your job is to organize it, not to rewrite it.

What to remove:
False starts, filler words, and repeated sentence beginnings.
Places where I said the same sentence twice in a row while finding it.
Verbal tics that carry no meaning.

What to keep, without exception:
My words. My sentence structures. My contractions. My rhythm, including the sentences that run long and the ones that are fragments.
My word choices, including the ones you would improve.
Anything that sounds like a digression. I will decide what is a digression.

What not to do:
Do not add transitions I did not say.
Do not smooth or vary my sentence openings.
Do not replace a plain word with a better one.
Do not fill anything in. Not one clause.

Markers, inline, in square brackets:
[GAP] where I clearly lost the thread or trailed off.
[SUPPORT] where I made a factual claim without support.
[PICK ONE] where I said the same thing twice in genuinely different ways. Keep both.
[CIRCLED] where I returned to an earlier point. Group those passages together and mark the group.

Output, in this order:
1. The cleaned transcript with markers.
2. A list of everything you removed, quoted, one per line.
3. Nothing else. No summary, no assessment of the chapter, no suggestions.

The same discipline applied to a draft you typed rather than spoke.

Read this draft and mark every place where the reasoning has a hole: a step skipped, a conclusion that does not follow, a term used before it is defined, or a claim that assumes something I have not said.

Mark them inline with [GAP: what is missing] and return my text otherwise unchanged, character for character. Do not fill any gap. Do not fix any sentence.

Quick, mechanical, and it will find more than you expect.

Go through this chapter and mark every sentence that makes a factual or empirical claim without support in the surrounding text. Insert [SUPPORT] immediately after each one and change nothing else.

Include claims I have phrased as common knowledge. At the end, list which ones you think a hostile reader would actually challenge, as opposed to which are merely unsourced.

Abstract passages are usually abstract because the author never had a case in mind.

Read this chapter and tell me the three places where the reader most needs a concrete example and does not get one. Quote the sentence each gap sits after and say what kind of example would land there: a case, a number, a piece of dialogue, or a before and after.

Do not invent examples. Do not write any. I have them and you do not.

Analogies fail at the edges, and the edge is where a smart reader is standing.

Here is an analogy I am using to explain [concept]: [paste].

Tell me where it breaks. Specifically: what does the analogy imply that is not true of the real thing, what important feature of the real thing has no counterpart in the analogy, and what wrong conclusion could a reader reasonably draw from it.

Then tell me whether the breakage is fatal or whether one sentence of acknowledgment fixes it. Do not propose a replacement analogy.

Do this once per major character and keep it in your bible. Chapter 14 has the axes in full.

Interview me about [character] until we have a voice specification on five axes. Ask about one axis at a time and do not move on until my answer is specific enough to be falsifiable.

The axes: diction, meaning the register and vocabulary they actually use. Rhythm, meaning sentence length, whether they interrupt, whether they finish. What they will not say, meaning the topic they deflect. Their verbal tic, or the deliberate absence of one. And what they are always doing with their hands or their body while they talk.

Push back on any answer that would also be true of a generic person. "Sarcastic" is not an answer. "Answers questions with questions when she is afraid" is.

At the end, give me the five axes as I described them, in my words, in a block I can paste into my story bible. Do not add traits I did not give you.

Not the character being asked what they want. The character being asked what they are avoiding.

Interview me about [character] as a hostile biographer would. You believe they are lying to themselves and your job is to find where.

Ask about the thing they claim to want, and then about what they actually spend their time doing. Ask what they have told nobody. Ask what their closest friend would say behind their back. Ask what they were like at nineteen and what they would say now about that person.

One question at a time. Do not accept an answer that flatters them. If I give you a virtue, ask what it costs the people around them.

The four part test, run as diagnosis rather than generation.

Here is what I have for [character]: [paste].

Fill in a table with four rows: want, meaning the conscious external goal. Need, meaning what they actually require to be whole. Ghost, meaning the wound in the past. Lie, meaning the false belief they hold about themselves or the world.

Use only what is in my material. Where I have not given you one, write [MISSING].

Then tell me: does the want contradict the need? Does the lie explain the want? If either answer is no, that is the problem with this character and you should say so plainly.

Two characters doing one job is the fiction equivalent of three chapters doing one job.

Here is my cast list with a paragraph on each. For every character, state in under ten words their structural function in the story: what they do to the protagonist or plot that nobody else does.

Then show me every pair whose functions overlap, and for each pair tell me what would break if I merged them into one person.

Do not propose new characters. Do not write anything about them beyond the function line.

For the character who appears in three scenes and needs to be a person anyway.

Interview me about [minor character] for exactly six questions. Your goal is to find the one concrete, specific, slightly odd detail that makes them real: an object, a habit, a phrase, a thing they carry, a thing they refuse.

Do not offer suggestions. Ask, and when I give you something generic, tell me it is generic and ask again.

Run it on a full draft, per character, when the manuscript has grown past what you can hold in your head.

Below is every scene containing [character], in order, extracted from my manuscript.

Build a table of their behavior: chapter, what they want in the scene, how they speak, and what they do under pressure.

Then flag: any place their diction shifts without a reason in the story, any place they act against the want or lie in my bible, and any place they know or do not know something inconsistent with earlier scenes.

Quote the specific line for each flag. Do not rewrite dialogue. Do not suggest replacement lines.

Useful and dangerous. Everything it produces is a note, never a line of manuscript. It is thinking out loud with a puppet, and the puppet is not your book.

You are going to answer as [character], based only on the specification below. Stay in role until I say stop.

I am going to ask you about the events of chapter [n] from your point of view.

Two rules. Nothing you say is canon. It is my raw material and I will decide what is true. And if my question requires a fact you were not given, say "not in the spec" rather than inventing it, because an invented fact will end up contradicting my book.

The workhorse. Run it per chapter, not per book, and paste only the bible entries the chapter touches.

Below are the relevant story bible entries, then chapter [n] of my manuscript.

Find every contradiction between them. Physical description, age, timeline, geography, relationships, capability, established rules of the world, previously stated facts.

Return a table: what the chapter says, what the bible says, chapter line quoted, and severity of high, medium or low.

Report only conflicts. Do not report things that are merely unmentioned. Do not fix anything and do not rewrite my text.

The most common continuity failure in assisted fiction is knowledge state, not eye color, and no tool tracks it.

From the manuscript below, build a knowledge state grid.

Rows: every named character. Columns: every secret, revelation or significant fact in the story. Cells: the chapter number in which that character learns that fact, or "never", or [UNCLEAR] if the text does not establish it.

Then, separately, list every scene where a character acts on, refers to, or reacts to information their row says they do not yet have. Quote the line.

Do this after the first draft. It is faster than making a calendar by hand and it finds the Tuesday that lasts four days.

Extract a timeline from this manuscript.

Columns: chapter, what happens, the elapsed time since the previous entry as the text states or implies it, and the running total from the story's day one.

Where the text gives no time signal, write [UNANCHORED]. Where two passages imply incompatible intervals, flag it as [CONFLICT] and quote both.
Do not smooth over anything. Do not guess. A conflict I know about is worth more than a timeline that looks tidy.

For the scene where somebody sits down twice and a coat leaves the room by itself.

Read this scene as a stage manager. Track every physical element: who is present, where each person is, what each is holding, doors, lights, weather, injuries, clothing.

List every physical impossibility, every object that appears without arriving, every object that disappears without being put down, and every person who acts while offstage.

Return a plain list with the line quoted. Do not fix the scene.

Two rules from Chapter 15 are built into this prompt, and both of them are counterintuitive until you have watched an entry misbehave.

Turn my notes on [character or place] into a story bible entry.

Two hard rules.
Physical description goes in a separate NOTES block, not in the main entry, because the main entry gets injected into every generation that mentions this name and repeated description is why assisted prose keeps re-mentioning someone's green eyes.
No spoilers anywhere in the entry. The model reads the whole entry regardless of where I am in the story, so anything about a later reveal will leak into an earlier chapter. Put later material in a separate block labeled LATE, which I will add manually when I get there.

Format: ENTRY, then NOTES, then LATE.
Use my facts only. Where my notes are silent, leave it out rather than inventing a plausible detail.

Run it over eight to twelve scenes at a time, never the whole manuscript at once, because accuracy falls off as the input grows. Every "no" row is a scene that needs a reason to exist.

For every scene in the manuscript below, fill one row of a table.

Columns: scene number, the value at stake in one or two words, the charge on that value at the opening as positive, negative or neutral, the charge at the close, and turned yes or no.

A scene turns if the charge at the close differs from the charge at the open. Neutral to neutral is a no.

Judge only what is on the page. Do not be generous. Do not explain your reasoning unless the answer is genuinely ambiguous, in which case add one line.
Do not suggest fixes for the no rows and do not rewrite any scene. You will pad them and I will cut them.

Complementary to the turn audit and quicker.

For each scene below, tell me: what the point of view character wants in this scene, what stands in the way, and what they leave with.

If any of the three is not on the page, write [ABSENT] rather than inferring it from the surrounding chapters. A scene with two absences is the finding, not a failure of the exercise.

The only reliable diagnostic for voice sameness in a cast, and the cheapest test in this book. Chapter 16 explains how to read the score, and the short version is to calibrate against a book you admire before you judge your own number. Strip the tags yourself before you paste, and do not tell it who is in the scene.

Below is a dialogue exchange with all speaker attributions, action beats and names removed. Each line is numbered.

There are [n] speakers. Assign every line to a speaker, labeled A, B, C.

For each line, give your confidence as high, medium or low, and in under ten words the specific linguistic evidence you used: vocabulary, sentence length, syntax, what they refuse to say, how they handle a question.

Where the only evidence available is the content of the line rather than the manner of it, mark it [CONTENT ONLY]. Those are the lines that prove nothing about voice.

Return the numbered assignments and nothing else. No commentary on the writing. No summary. Do not attempt to improve any line.

After I tell you the correct answers, report your accuracy as a percentage and list which speakers you confused with which.

For the scene where everyone is saying exactly what they mean, which is the flatness tell nobody names.

Take this dialogue scene and build a three column table: the line as written, what the speaker literally says, and what the speaker actually wants from the other person in that moment.

Where columns two and three are the same, mark the row [ON THE NOSE].

Return the table only. Do not rewrite any line and do not propose subtext I have not implied somewhere in the scene.

The dialogue pass that improves almost every scene.

In this dialogue, identify every line that could be deleted without the reader losing information, character, or tension. Include greetings, acknowledgments, restatements of what just happened, and lines that exist only to hand the other person a cue.

Quote each line and give the one word reason. Do not delete them yourself and do not rewrite what remains.

Before you write it, not after.

Here is my brief for the next scene: [paste].

Ask me the questions I have not answered. What does each present character want, why is this scene here rather than summarized, what changes by the end, why does it start where it starts, and what would break in the book if I cut it.

One question at a time. If my answer to the last one is "nothing much," say so.

Run it after the scene exists, with the brief pasted above the draft.

Above is the brief for this scene. Below is the scene I wrote.

Tell me, point by point, which elements of the brief made it onto the page, which are present but thin, and which are absent. Quote evidence for each.

Then tell me what is in the scene that is not in the brief, and whether it is a discovery worth keeping or a drift worth cutting. Do not revise the scene.

Start here. It takes ten seconds and it is right more often than it has any right to be.

What is the weakest paragraph in this chapter, and why is it the weakest? Answer in three sentences. Do not rewrite it.

Not whether they like it. Where they put it down.

Read this as someone who is skeptical of the whole premise and is looking for a reason to stop.

Give me the three most likely paragraphs at which they close the book, in order, with the exact sentence that does it and what specifically triggers it: a claim they will not accept, a stretch of nothing happening, a tone they resent, or a demand on their patience I have not earned.

Do not fix anything.

Ask it of chapters you love. Those are the ones carrying sentiment instead of work.

Make the strongest case that this chapter should be cut entirely from the book.

Consider: its job is done elsewhere, it slows the book at a bad moment, it is the author's interest rather than the reader's, and its content would fit in three paragraphs in a neighboring chapter.

Then tell me what the book would lose. Be specific about what the reader would not get.
No recommendation and no encouragement. I will decide.

The single most valuable question in this book. Ask it of every chapter you write, and expect it to name the thing you have been avoiding, which is usually the reason the chapter is not working.

A reader who disagrees with what this chapter argues: what is their best objection?

Not their most common objection and not their most emotional one. The one that is hardest for me to answer, stated the way an intelligent person who has thought about this would state it, at full strength and without hedging.

Then tell me whether the chapter addresses it, ignores it, or works around it while pretending to address it.

The one that finds the sentence you have now written four times.

Across the chapters below, find repetition. Three kinds, listed separately.

Repeated ideas: the same point made in more than one place. Quote each instance and tell me which is the strongest.
Repeated phrasings: distinctive phrases, images or constructions I have reused. Give a count.
Repeated structures: openings, transitions or paragraph shapes I have fallen into.

Do not include ordinary words and do not treat deliberate refrain as an error. Change nothing.

Chapter 20 has the full check and the repairs. This is the version you run at speed.

Read this passage against the Voiceprint above and report on flatness, using measurement rather than impression where you can.

Report: contractions per thousand words against my baseline, average sentence length and its variance against my baseline, count of sentences opening with the same construction, count of hedges such as somewhat, rather, quite and arguably, count of abstract nouns where a concrete one is available, and any place a specific detail has been replaced by a category.

Then name the three sentences that read least like the Voiceprint and say why for each.

Do not repair anything. Do not offer a rewritten version even as an illustration.

The last structural pass before you stop revising.

Go through this chapter and list everything I am asking the reader to accept.

For each, mark it as: supported in this chapter, supported elsewhere in the book and cross referenced, supported elsewhere but not cross referenced, or unsupported.

For the unsupported ones, tell me what kind of support would be sufficient: evidence, an example, a concession, or a narrower claim. Do not write the support.

Describe the mess you have and the thing you want. Paste a real sample.

Write me a script that converts [input format] to [output format]. I am on [operating system] and I would prefer [language, or say "whatever is simplest to run"].

Here is a representative sample of the input, including the ugly parts: [paste].

Requirements: it must not modify the original files, it must tell me what it did, and it must fail loudly rather than silently skipping anything it does not understand.
Explain how to run it as if I have never used a terminal.

For the EPUB checker output that reads like a threat.

Here is the output from an EPUB validator on my file: [paste].

For each error, in plain English: what it means, whether it will actually stop the file from being accepted or is only a warning, which file and line to look at, and the smallest change that fixes it.

Sort by whether it blocks publication. Do not tell me to use a different tool.

Alt text is accessibility metadata, not prose, and it is exactly the kind of small necessary job that never gets done by hand.

For each image I paste, write alt text of about 125 characters, which is a screen reader convention rather than a hard limit, that describes what the image conveys in the context of a book about [subject].

Describe function, not decoration. If the image contains text, include the text. If the image is purely decorative, say so and mark it as such.
Return a numbered list matching my image numbers. No commentary.

Your numbers, its plotting. Never ask it to supply the numbers.

Here is a table of data I collected: [paste].

Write code to produce a clean chart of it suitable for print at 300 dpi in grayscale, sized for a 6 by 9 inch page with 0.75 inch margins.

Use only the numbers I gave you. Do not add data points, do not extrapolate, do not smooth. If the data does not support the chart type I asked for, tell me instead of drawing it.

For a cover designer or an image tool. The brief is the deliverable, not the picture.

Interview me until you can write an art direction brief for my book cover, then write it.

Ask about: the genre shelf it must sit on, three comparable covers and what specifically I like in each, the emotional register, what must be legible at thumbnail size, what must not appear, and any object or image that is central to the book.

Output the brief in under 250 words as instructions for a human designer: concept, mood, composition, color direction, typography direction, and hard constraints. No adjectives that could describe any book.

Quantity first, judgment second. You are looking for one.

Here is my book: [one paragraph]. Reader: [who].

Write fifteen one sentence hooks for the top of the product description. Vary the approach: the problem, the counterintuitive claim, the reader's current situation, the question, the promise.

No colons. No rhetorical questions in more than three of them. No "in the current world." Number them.

The structure matters more than the phrasing.

Write my Amazon product description in this structure: a hook line, then two or three short paragraphs of what the reader gets, then a bulleted list of three to six specific things covered, then one line of why I am the person writing it, then a one line close.

Book details: [paste]. Reader: [paste]. Comparable titles: [paste].

Constraints: under 300 words, no income claims of any kind, no promises about results, seventh grade reading level, no adjective that could describe any other book in the category. Give me three variants with genuinely different hooks, not three phrasings of one.

Read the built-in instruction in this one carefully, because it is the point. A model has no access to Amazon's query logs and cannot know search volume. Asked for it anyway, it produces a confident invention, and that is the most common way authors waste money on assisted metadata.

Generate candidate keyword phrases for my book. Details: [paste].

Give me 40 phrases a real reader might type into an Amazon search box, grouped by intent: problem searches, solution searches, comparison searches, and genre or shelf searches.

Explicit instruction: do not estimate search volume, competition, difficulty, or ranking potential for any phrase. You have no access to Amazon's search data and any number you produce would be invented. Do not label anything as high or low volume. Do not sort by estimated traffic. If you are inclined to add that information, add nothing instead.

This is a seed list. I will check every phrase in Amazon's own search bar and in a real keyword tool.

The model helps you read the tree. It does not know what is selling in it.

Here is my book: [paste]. Here are the categories I am considering: [paste the full category paths].

For each, tell me: what a reader browsing that category is looking for, whether my book delivers that, and what kind of book would beat mine there.

Then rank them by fit to the book, not by anything about sales, competition or rank, which you cannot know and must not estimate. Flag any category where my book would be the odd one out.

The warning is the important half. A model writes a competent email that reads exactly like every other competent email.

Draft a short email to [person or outlet] about my book. Details: [paste]. Why they specifically might care: [paste].

Under 150 words. No flattery in the opening. State what I want in the first three lines. One specific reason it is relevant to them and their audience.

Then, separately, list the three sentences in your draft that are generic enough to appear in anyone's pitch. I will replace those with something only I could write, and if I do not, this email is not worth sending.

Volume is the point. You will test them and most will lose.

Write 40 ad headlines and 20 body lines for [platform], for my book: [paste].

Constraints: [character limit] per headline. No income claims, no urgency language, no "bestselling" unless I tell you it is true. A third should lead with the problem, a third with the reader's identity, a third with a specific detail from inside the book that I have given you.

Number them so I can report back which ones worked.

One piece of source material, a month of posts, all of it drawn from words you already wrote.

Here is a chapter from my book, and below it a transcript of me talking for ten minutes about what writing it was actually like. Turn the pair into a month of social content, and use the transcript for the voice.

Produce: 12 short posts, 4 longer posts of about 200 words, 6 single sentence observations, and 3 questions to ask my audience.

Every item must be traceable to something specific in the chapter. After each one, cite the section it came from. Do not introduce claims, statistics or examples that are not in my text.
No hashtag stuffing. No opening line that is a rhetorical question in more than three items total.

For when the post is written and the first line is not landing.

Here is a post. Give me eight alternative opening lines: two blunt, two specific and concrete, two that state a number or detail from the post, two that name the reader's situation.

Keep the body unchanged. No questions.

The thing people trade an email address for. It should come out of the book, not beside it.

From my book's material below, propose three reader magnets: a short, genuinely useful thing someone would give an email address for.

For each: title, what problem it solves in one line, the format, roughly how long, and which part of the book it comes from.

It must be useful on its own and it must make the book more likely, not less. Do not propose a sample chapter. Outline only, in bullets. I will write it.

Three emails, written once, working for years. Chapter 29 argues for three rather than ten, on the grounds that three gets written.

Draft a three email welcome sequence for people who join my list after downloading [magnet]. My book: [paste]. My reader: [paste].

Email one delivers the thing. Email two says who I am and how often I will write. Email three gives something useful on its own and mentions the book once, plainly.

Under 200 words each. Plain text style, no headers or graphics. No income claims, no urgency, no fake scarcity. Leave [PERSONAL] markers where a specific story of mine would go, and do not invent the story.

The prompt that keeps a list alive when you have nothing prepared.

Below are my rough notes from this week: fragments, links, half thoughts.

Organize them into a newsletter of about 500 words with one main idea and two short items. Use my phrasings where I made them. Mark [THIN] where an item needs a sentence from me to work, rather than filling it in.

Suggest three subject lines. No preamble about what a week it has been.