How to Write a Book With AI: An Eight-Stage Workflow From Notes to Chapters

Most experts who try writing a book with AI start in the same place. They open a chat window, type the book idea and ask for an outline. Twenty seconds later they have twelve chapter titles that could belong to anyone in their field.

The material that would make the book theirs is somewhere else: a folder of keynote decks, three years of newsletter issues, de-identified session notes, a webinar recording with a good Q&A. The workflow below starts there and keeps one rule throughout. You decide what the book argues. The AI sorts, drafts and checks, and only from what you gave it.

If you have not yet decided whether a general chat tool is enough, read whether ChatGPT or Claude can write a business book first. This article assumes you are going ahead.

The eight stages on one page

Each stage produces one thing you approve before the next begins. That is what stops a small outline problem becoming twelve chapters of rework.

Stage	You do	AI does	You approve	Failure to watch for
1. Gather and de-identify	Choose files, remove client details	Lists topics and where they appear	A source inventory	A client recognisable in a story
2. Positions	Write 8 to 15 claims you would defend	Asks sharpening questions, nothing more	A positions page	AI-suggested opinions slipping in
3. Outline	Compare, choose, reorder	Proposes structures from your material	A chapter list	Chapters with no source behind them
4. Chapter briefs	Set each chapter’s job and sources	Drafts the brief with source tags	One brief per chapter	Gaps hidden in vague wording
5. Drafting	Supply notes per section, review	Drafts one section at a time	Section drafts	Invented examples
6. Claim check	Decide every flagged claim	Matches each claim to a quote	A claims list	“Checked” with no quote attached
7. Voice pass	Read aloud, mark lines	Rewrites only marked lines	A revised chapter	Polish replacing your phrasing
8. Whole-book pass	Read it end to end	Runs narrow consistency checks	A manuscript for an editor	Trusting one giant prompt

In stage 2 the AI does almost nothing, and in stage 8 it only answers narrow questions. That is deliberate.

Before the AI writes a word: stages 1 and 2

Stage 1: Gather the material and remove who it is about

Pick the documents that carry your thinking, not every file you own. Remove client names, company names and any detail that would let a former client recognise someone, before anything is uploaded. The full audit of which files earn a place is in writing a book with AI from documents you already have. For the privacy side, see protecting your ideas and clients when writing with AI.

The AI’s job at this stage is an inventory, not a summary. Anthropic’s prompting guide recommends putting long documents near the top of the prompt, above your instructions, and wrapping each one in tags with its source name. It reports that questions placed at the end “can improve response quality by up to 30 percent in tests”, especially with several documents (Anthropic).

[Your documents go above this line, each wrapped in <document> tags with its file name.]

List the topics these documents cover. For each topic, give the file names
and one short exact quote showing where it appears. Do not summarise or add
ideas that are not in the documents. If a topic appears in only one file,
mark it THIN.

The failure to watch for is a client who is still recognisable: a changed name does not help if the industry, city and headcount remain.

Stage 2: Write your positions yourself

Before any outline, write one page of positions: eight to fifteen sentences, each a claim you would defend in a client meeting. “A leadership offsite without a named owner for every decision is a wasted day” is a position. “Offsites can be valuable” is not.

This is the stage we would never hand to AI, even partly. Positions are what the reader pays for, and a model asked to suggest them produces the common view of your field, which your book should be disagreeing with. If the page is hard to write, extracting your knowledge before you write covers ways to get it out of your head.

AI can still help, as an interviewer:

Here are the positions I want my book to argue. Do not add positions,
rewrite them or suggest new ones. For each, ask me one question that would
make it sharper: a likely objection, a case where it fails, or what I would
say to a client who disagrees. Ask one question at a time and wait for my
answer.

Watch for the interviewer “helping” by offering a neater phrasing that you accept without noticing. Save your answers: they are raw material for later chapters.

Stage 3: Using AI to create a book outline from your material

Now give the AI the positions page and the stage 1 inventory, and ask for two structures. One outline is easy to accept; two force you to judge.

[Positions page and topic inventory above this line.]

Using only the positions page and the inventory, propose two different
chapter structures for a nonfiction book for [describe the reader]. Each
chapter must argue at least one of my positions and list the source files it
would draw on. If a chapter would need material that is not in the
inventory, include it but label it GAP and say what is missing. Do not
invent topics, examples or frameworks to fill the structure.

Then reorder by what the reader needs first, which is rarely the order your material was created in, and cut any GAP chapter you cannot fill yourself. If more than two chapters are GAP, go back to stage 1.

The failure here is the generic chapter: “The Importance of Mindset” with no files behind it. For the structure of the outline itself, use our business book outline template, or the main-site guide to outlining an authority book.

Stage 4: Turn each chapter into a brief with source tags

An outline names each chapter. A brief says what it must do and where every section’s material comes from. Most people skip it; it is what makes stage 5 checkable.

Give each source passage a short tag, such as [S4: March webinar, 12:30 to 18:00] or [S9: 2021 offsite deck, slides 12 to 19], and keep the tag list in one document.

For chapter [n], draft a one-page brief with: the question the reader
arrives with, what they should be able to do after reading it, the positions
it argues, and the sections in order. Under each section, list the source
passages it will use by tag. If a section has no source, write NO SOURCE.
Do not fill it.

Every NO SOURCE gets a decision now, not during drafting: supply it (record yourself answering the question), cut it, or research it and cite a real publication. The chapter-level pattern is in how to write a business book chapter.

Stage 5: Using AI to write book chapters from notes, one section at a time

This is where the workflow holds or quietly fails. Draft sections, not chapters: roughly 600 to 1,200 words per request, short enough to check every sentence against its sources in one read. Whole-chapter requests are a false economy, because nobody checks 5,000 words as carefully as 800.

Paste only that section’s material. Two of Anthropic’s documented techniques for reducing invented content apply directly: explicitly allow the model to say it does not know, and instruct it to use only the documents provided rather than its general knowledge (Anthropic).

<sources>[the tagged passages for this section]</sources>
<brief>[the section brief]</brief>
<voice>[your voice notes and two paragraphs you wrote yourself]</voice>

Draft section [n] of chapter [n], about [length] words, using only the
sources above, not your general knowledge. Keep my examples, phrases and
order of argument where they work. After each factual sentence, add the tag
of the source it came from. Where the argument needs an example, figure or
client result that is not in the sources, write [NEEDS: what is missing]
and continue. Do not create one.

A hypothetical example. A leadership coach is drafting a section of her chapter on why offsites fail, from eleven minutes of webinar transcript, two pages of de-identified session notes and one slide. The draft comes back fluent, tagged, and with one [NEEDS: a result from a team that assigned decision owners]. She has such a result but not the client’s permission, so she writes a composite, labels it as one in the text, and moves on. That is a judgement no prompt could make for her.

Watch for the plausible story with no tag, or a tag pointing to a passage that says something else. Spoken material needs its own repairs; see turning talks and sessions into chapters.

Stages 6 and 7: check the claims, then the voice

Claims come before voice, because a voice edit rewrites exactly the sentences you need to trace back to a source. Polish first and you will be checking prose that no longer matches its tag.

Stage 6: Claim check against the sources

This prompt adapts the “verify with citations” technique from Anthropic’s guide: for each claim, find a supporting quote, and retract the claim if none exists.

[Drafted section and its sources above this line.]

List every factual claim in the section: numbers, client results, quotes,
dates, named methods, anything a reader could check. For each one, copy the
exact quote from the sources that supports it. If you cannot find one, mark
the claim UNSUPPORTED. Do not judge whether a claim is true from general
knowledge.

Run it in a fresh session, so the checker is not grading its own homework. Then read every UNSUPPORTED line and spot-check a few supported ones; Anthropic is plain that these techniques reduce invented content but “don’t eliminate them entirely”. The prevention side, including an approved claims list, is in how to stop AI inventing facts in your book, and the main site has a fact-checking guide for AI-written nonfiction.

Stage 7: Voice pass on marked lines only

Read the checked section aloud and mark with [V] every sentence you would never say to a client. Then ask for targeted changes, never a global “make this sound more like me”, which pushes the whole section towards smoother, more generic prose.

I have marked sentences with [V] that do not sound like me. Using my voice
notes and sample paragraphs, rewrite only the marked sentences. Do not
change facts, source tags, unmarked sentences or the order of the argument.
Show each change as before and after.

Building the voice notes is its own job; see training AI on your writing style.

Stage 8: The whole-book pass, done in pieces

The tempting move is to paste the full manuscript into one chat and ask whether it is consistent. Anthropic’s documentation explains why that disappoints: “more context isn’t automatically better. As token count grows, accuracy and recall degrade, a phenomenon known as context rot” (Anthropic). The answer you get will sound complete and will not be.

So split the pass. You read the whole book in one or two sittings; this step needs a single mind holding every chapter. The AI runs narrow checks, one question at a time: every use of one framework term, every repeated story, every promise the introduction makes.

[Glossary and manuscript above this line.]

For the term [framework stage name], list every chapter and sentence where
it appears, and every place a different name seems to refer to the same
thing. Quote each sentence. Do nothing else.

Repeat per term and per story. It is dull, and it catches the drift that loses readers in chapter nine. For voice drift, see keeping one voice across a book. Then the manuscript goes to a human editor; using AI to edit a business book covers what AI can add there.

Where to run it: a chat tool or guided software

A general chat tool can run all eight stages. As of November 2026, Claude’s Projects give you “self-contained workspaces with their own chat histories and knowledge bases”, plus project instructions; free accounts can create up to five, and on paid plans the knowledge base switches to retrieval as it nears context limits, “to expand capacity by up to 10x” (Claude Help Center). Keep the positions page, tag list, glossary and voice notes there. Retrieval pulls only the passages judged relevant, so in stage 5 we would still paste each section’s sources directly. Other chat tools offer similar workspaces; we have not verified their current limits. Either way, the tags, claims list and drafting log are yours to maintain.

WritebookAI is one option that builds several of these stages in. Its source analysis reads PDF, Word and text files or pasted notes and shows what the material covers and what is missing; a verification interview resolves conflicts and gaps; and a strategy and chapter plan with beat sheets comes before any drafting. Chapters are drafted from your material with claim checks: when something is missing, such as a client result or an unsourced statistic, the chapter leaves it out and asks you rather than inventing it, and you approve every chapter. A Writing DNA voice profile is built from your own samples and tested on a fresh topic before it guides a chapter, and there is a whole-book manuscript review at the end. It does not write your positions or replace your fact-checking and a final copyedit, and it is not done-for-you ghostwriting. It is the wrong choice for fiction or for anyone after the cheapest AI writer. Details at writebookai.com.

Your first two hours

Tonight, write the positions page. It is the only stage no tool can do for you, and everything after it depends on it. Ten sentences is enough to start.

Tomorrow, de-identify five documents and run the stage 1 inventory prompt. Note which positions have three files behind them and which have none. That gap list is the most honest outline you will see this month, and it tells you whether the book is ready to draft or needs another week of gathering.