How to Write a Book With AI Using Documents You Already Have

Somewhere on your laptop is a folder that could become a book. A keynote deck you have given forty times. Three years of newsletters. Two proposals that explain your method better than your website does.

The obvious move is to drop the whole folder into an AI tool and type “turn this into a book”. What comes back is long, smooth and oddly empty: your headings, someone else’s sentences, and a client story you do not remember telling.

The files were fine. The missing piece is the step before the upload.

Why “upload everything and ask for a book” goes wrong

Three separate failures hide in that one request.

The model cannot hold it all at once. Anthropic’s help centre says the standard context window on paid Claude plans is 200K tokens, “about 500 pages of text”, with some newer models listed at up to 1M tokens (Claude Help Center, as of November 2026). A decade of transcripts, newsletters and proposals can pass that sooner than you would expect.

When it cannot hold it all, it searches instead. Claude Projects switch to what Anthropic calls RAG mode “when your project approaches or exceeds the context window limits”, after which Claude “searches and retrieves only the most relevant information needed to answer your questions” (Claude Help Center). Retrieval works on fragments: OpenAI’s developer documentation, for example, says its file-search system splits files by default into 800-token chunks (OpenAI), a few paragraphs each. A request as broad as “write my book” gives the search little to aim at, so the model sees whichever paragraphs scored highest.

Even what it can see, it does not use evenly. Liu and colleagues found that model performance “is often highest when relevant information occurs at the beginning or end of the input context” and degrades for material in the middle (Liu et al., TACL). That study dates from 2023 and models have improved, but a 2025 Chroma report testing 18 newer models still found that “model performance consistently degrades with increasing input length” (Chroma). Chroma sells retrieval software, so weigh that.

So the model has partial sight of your material, no idea which parts are load-bearing, and an order to produce a whole book. Where your files are silent, it writes something plausible. That is a main reason AI-written books sound generic, and why we think “upload less, better chosen” beats “upload everything” almost every time.

Passes 1 to 3: choose, clean and de-identify before the AI sees anything

Pass 1: Choose documents by what they prove, not how many pages they fill

Ask of every file: what does this give the book that nothing else does? Four things count. A story (something that happened). A method (a framework or sequence you teach). Proof (a result, figure or source). Your voice (how you actually phrase things).

Document	Usually good for	Watch out for	Our verdict
Keynote and workshop decks	Method, teaching order, chapter-sized ideas	Slides without notes are headlines with no argument	Use only with notes or a talk recording
Talk, webinar and podcast transcripts	Stories, spoken voice, audience questions	Filler and repetition	High value once cleaned
Newsletters and long posts	Voice, opinions, ideas that resonated	Topical hooks that date quickly	Use the best twenty, not all of them
Proposals and programme outlines	Your method as you sell it	Client names, fees, contract terms	Use after redaction
Course notes and workbooks	Exercises and tools	Written for a learner mid-course	Mine for tools, rewrite the framing
Short social posts	Sharp one-liners	Fragments with no argument	Mostly overrated as source material
Raw session notes	Recurring client problems	Confidential by default	Keep out of AI tools; summarise patterns by hand

If your material is scattered across years and formats, sort it first; our guide to organising years of experience into book chapters covers that inventory. Here we assume you roughly know what the book is about and are deciding what the AI should see.

Pass 2: Clean the files so the AI reads your thinking, not your formatting

Cleaning is dull, and much of the quality comes from it. Aim for one text or Word file per source, opening with a line saying what it is, its date, and whether you still believe it.

• Remove duplicates and old versions. If the 2019 and 2024 decks disagree, keep the one that reflects how you teach now. Contradictory sources invite the model to blend them.
• Convert decks properly. Export speaker notes with the slide text, or use the transcript of the talk instead.
• Strip boilerplate. Newsletter footers, “book a call” blocks and slide numbers all get read as content.
• Tidy transcripts lightly. Delete the ums and tangents, but do not rewrite your phrasing; that is the voice you will want the AI to match (see training AI on your writing style for a book).
• Mark what is out of date. A note such as “I no longer recommend step 3” saves you from a chapter built on an abandoned method.

Pass 3: De-identify before anything leaves your laptop

Your obligations do not change because the tool is new. The ICF Code of Ethics (in effect since 1 April 2025) asks coaches to “maintain the strictest level of confidentiality with all parties involved”, and Standard 2.5 extends their obligations to the technology they use, naming AI explicitly. Consultants usually have the same duty in their contracts and NDAs.

Replace each removal with a tag such as [CLIENT A, SaaS, 40 staff] so the story still works. Whether a disguised story can appear in print is a separate decision; the main-site guide to using client stories in a business book covers consent and disguise.

Pass 4: Map coverage against your chapter plan

This step separates a book from a compilation. Take a rough chapter plan and build a grid: chapters down the side, the four kinds of evidence across the top. Fill each cell with the source that supplies it, or leave it empty.

A worked example (hypothetical)

Imagine an operations consultant, Tomás (invented for this example), with eleven years helping founder-led companies hand off daily management. After Passes 1 to 3 he has 14 cleaned files: four transcripts, two decks with notes, five newsletters, two redacted proposals and a workbook. His first five rows:

Chapter	Story	Method	Proof	Voice	Status
1. The founder is the bottleneck	Keynote transcript (own company)	—	Newsletter 12 cites a survey	Strong	Proof needs tracing
2. Decisions, not tasks, get delegated	Proposal B (Client A)	Workbook: decision-rights map	—	Strong	No proof yet
3. Writing the operating rhythm	—	Deck 2 with notes	—	Thin	Empty story
4. Hiring the first operator	Podcast transcript	—	—	Medium	Thin
5. The founder’s new job	Newsletter 30	Deck 1	Proposal A outcome (needs consent)	Strong	Consent pending

Chapter 3 has a method but no story; draft it now and the AI will invent a founder who struggled with weekly meetings, convincingly. The survey in chapter 1 was quoted in a newsletter, never opened; it needs tracing to its publisher or it comes out. Chapter 4 rests on one podcast anecdote.

Fourteen files sounded like plenty. The grid shows three chapters close and two that need Tomás himself, not more documents.

Pass 5: Decide how each gap gets filled, before drafting starts

Every empty or thin cell gets one of four decisions, written in the grid:

1. Talk it out. Record yourself telling the missing story or explaining the step, transcribe it, and add it as a source. For most experts this is the fastest fix; see turning talks, webinars and sessions into a book and the main site’s guide to turning podcasts and coaching calls into a book.
2. Find the original. Trace every figure to the publisher that produced it. If you cannot, cut it.
3. Ask for permission. For client results you want to name, ask now, not the week before the manuscript is due.
4. Cut or merge the chapter. A chapter with no story, method or proof is often an idea you like rather than one you can teach. We would cut it before letting an AI write it.

What you never do is mark the gap “AI to write”. The model was not in the meeting you are thinking of.

Keeping the AI inside your material when you draft

Once the grid is filled, draft one chapter at a time from that chapter’s sources plus your voice samples. Smaller inputs avoid most of the problems above and keep every sentence traceable.

Anthropic’s guidance on reducing hallucinations lists the techniques that work here: let the model say it does not know, have it extract word-for-word quotes before writing, restrict it to the provided documents, and make it find a supporting quote for each claim afterwards, removing and marking any claim it cannot support (Anthropic). The same page notes these techniques reduce hallucinations but “don’t eliminate them entirely”. As a chapter brief:

Draft Chapter 3 using ONLY the attached sources (S1–S4) and my voice samples.
1. First, list the exact quotes from S1–S4 you will rely on, with source labels.
2. Write the chapter. After every story, figure or client result, add [S#].
3. If the chapter needs a story, example or number that is not in the sources,
   do not write one. Insert [STORY NEEDED: what it should show] or
   [SOURCE NEEDED: the claim] and carry on.
4. After drafting, check each [S#] against the quote list. Remove any claim
   you cannot match and mark the gap with [ ].

The placeholders are the point. A draft full of brackets looks worse than a fluent one and is far more useful: every bracket is a question only you can answer. We go further on stopping AI from inventing facts in your book, and the main site has a checklist for fact-checking AI-written nonfiction.

General AI tools and guided book software: what they do with your files

You can run this whole method in a general assistant. Here is how three common ones handle uploads, per their own help pages as of November 2026.

• Claude Projects keep uploaded documents and project instructions together, use the full context while it fits, and switch to retrieval when it does not; Anthropic says RAG mode, on paid plans, can hold “up to 10x more content” (Claude Help Center). Good for chapter-by-chapter drafting with a standing brief.
• ChatGPT Projects also group files, chats and instructions (OpenAI Help Center). The number of files you can add depends on your plan and changes often, so we give no figures; check the help centre before planning around a limit.
• Google’s NotebookLM, now Gemini Notebook, accepts Word, PDF, text and PowerPoint files among others, up to 500,000 words or 200MB per source and 50 sources on the free tier, and “uses the sources you upload to answer your questions” (Google). Strong for interrogating sources; less suited to drafting a long book in one voice.

All three are capable, and none will audit your material for you. The coverage grid, gap decisions and “do not invent” rules stay your job, in every session. For more, see can ChatGPT write a business book and the main-site comparison of ChatGPT and an authority book platform.

Is there a tool that does the audit for you?

Partly. WritebookAI is one option built around this step: you upload PDF, Word or text files or paste notes, and its free source analysis shows what the material covers and what is missing. A verification interview then asks you about conflicts and gaps before you approve a book strategy and chapter plan. When a chapter needs something your material does not contain, such as a client result or an unsourced statistic, the draft leaves it out and asks you rather than inventing it, and you approve every chapter. It will not decide which documents you may share, de-identify them for you, or vouch for a fact you supply, and it is not a done-for-you service. If you would rather hand the files to a person, a ghostwriter suits you better; see writing a business book without a ghostwriter. Read the privacy policy of this or any tool before uploading client-derived material.

Mistakes that waste good material

• Uploading the archive instead of the audit. More files means more retrieval noise, not more book.
• Treating old posts as chapters. A 900-word post is evidence for a chapter, not the chapter.
• Trusting a figure because you once quoted it. Your newsletter is not the source.
• Deleting the placeholders to make the draft look finished. Answer them instead.

What to do this afternoon

Pick your five strongest documents, not your fifty. Clean and de-identify them using the checklist above. Sketch a rough chapter list and build the coverage grid in a spreadsheet. The empty cells are your real to-do list, and they say more about whether the book is ready than any draft could.

When most rows have a story, a method and some proof, draft chapter one from its own sources. More on tools sits under AI book-writing software.