How to write a book with AI: from source notes to a checked draft.

What AI can help with when writing a book

Give AI a small editorial assignment: find gaps in this outline, suggest three ways to explain this verified idea, turn these interview notes into a section draft, or test whether these instructions omit a step. Small assignments make the input, assumptions, and output inspectable.

The author still owns the difficult decisions: which reader deserves the book, which evidence is adequate, which experience is honestly theirs, what uncertainty must remain visible, and what is ready to carry their name.

Three AI prompts for an outline, draft, and revision

Replace the bracketed fields and attach your own material. These are reusable instructions, not benchmarked promises of output quality. Run them separately so you can inspect the outline before requesting prose.

1. Find overlap in the outline

Reader: [specific reader and starting knowledge].
Book outcome: [observable result].
Outline: [paste chapters].
For each chapter, state the question it answers and what the reader produces. Flag duplicate questions and missing prerequisites. Suggest a revised order and explain each move. Do not draft chapters yet.

2. Draft from supplied evidence

Draft [section] for [reader]. Answer only [question].
Use these notes: [source IDs and relevant excerpts].
Include this example: [your example, labeled real or hypothetical].
Do not invent facts, statistics, quotations, or references. Mark unsupported claims [VERIFY]. End with [specific reader action]. List unresolved questions separately from the prose.

3. Review without rewriting everything

Compare this draft with the section brief and supplied sources. Quote the passage containing each problem and classify it: unsupported claim, repetition, missing step, or voice mismatch. Propose the smallest useful correction. Preserve names, numbers, quotations, and source IDs unless you flag a specific discrepancy. Return the diagnosis before a rewrite.

After each response, check that the requested example survived, open the sources behind factual claims, and reject invented details even when they sound plausible. If the model loses context across chapters, supply a short record of established facts and the previous chapter's conclusion with the next assignment.

Choose a task the supplied material can support

With the volunteer notes above, AI can organize the observations, propose section headings and draft alternative instructions. It cannot know whether the map was implemented or how newcomers felt. Those questions require additional records or interviews.

For an outline, ask which reader question each chapter answers. For a factual paragraph, supply the relevant notes and check the resulting claims. For a rewrite, identify the defect and preserve accepted facts. Use a different assignment when the missing ingredient changes; an instruction to “write more confidently” cannot replace research.

Worked example: write from a source packet without adding facts

Here is a small invented source packet for a volunteer-group handbook. It demonstrates the editing method; it is not research about real volunteers. In your book, use records you have permission to share.

A draft that sounds plausible but exceeds the notes

“Confusing directions caused new volunteers to arrive late. After introducing a welcome map, the organization improved attendance and helped everyone feel included.”

There are four additions: causation, lateness, an implemented map and an improved result. None is supported by N1–N3. The last phrase also claims an emotional response that nobody reported. Replacing “improved” with “may improve” would still leave the invented event in the paragraph.

A version the packet can support

“At one Saturday session, two new volunteers asked where the sign-in sheet was. The printed welcome sheet also named a different meeting point from the room used that day. Before the next session, the coordinator proposed adding a map and checking the instructions with a newcomer.”

This version reports the observed problem and proposed next step. A following exercise can ask readers to check their own arrival instructions, but it should be presented as an exercise, not as a proven intervention. During review, attach N1 to the first sentence and N2–N3 to the following sentences. Remove these internal IDs from the reading edition only after replacing them with the attribution or source notes your book needs.

Label notes by what they can support

For the volunteer example, keep N1, N2 and N3 separate even though they concern the same session. Combining them into a summary such as “The welcome process was fixed” would erase the difference between an observation, a mismatch and a proposal before the drafting tool even starts.

• Original: retain the interview, document or observation note and its date. A generated summary is a working aid, not a replacement for the original.
• Source location: add a page, heading, timestamp or record identifier so another person can find the relevant passage.
• Supported use: state the narrow claim it permits. N2 supports a mismatch between written directions and the actual meeting room; it does not establish an attendance change.
• Restriction or open question: record permission limits, a disputed detail or something that still needs confirmation.

For a book drawing on published research, also record the population, time period and meaning of key measurements. A result about one group or setting should not become a universal rule during summarization. Verify current platform instructions from the platform's own documentation close to publication.

Move from a checked chapter to a complete draft

1. Approve a representative section. Use one that contains a source-based claim, an example and an instruction. Check all three against the brief.
2. Save the accepted decisions. Record terms, facts, unresolved questions and what the next chapter must add. The handoff example below shows the fields.
3. Draft the next chapter with its own source packet. Include the shared decisions, but do not assume that the first chapter's evidence supports a new claim.
4. Review the joins. Read the ending of one chapter and the opening of the next. Remove repeated lessons and repair any new assumption that contradicts the approved material.

Keep a correction list for the whole manuscript. If you change a definition or an example midway through, locate its earlier and later uses before treating the draft as complete. For the final pass, use the editing walkthrough and claim log.

A complete prompt for the volunteer-handbook example

Write a 180–250 word section called "Check the first five minutes" for a volunteer-group coordinator.
The reader needs to test whether a newcomer can find the meeting point and sign in.
Use only notes N1, N2 and N3 supplied above. Label the case as an example. Do not add attendance numbers, motives, quotations or an outcome for the proposed map.
Start with what was observed. Then provide a three-step exercise the reader can try at their next session. Clearly distinguish the exercise from events already observed.
After the prose, list each factual sentence and its supporting note ID. Put any missing information in a separate QUESTIONS section. Do not resolve missing information by guessing.

The length range keeps the assignment small enough to review; it is not a quality target. If the response spends half its space explaining why welcoming people matters, cut that introduction and retain the observation and exercise.

What to do when the response fails

If it invents a result, identify the exact sentence and source boundary: “N3 is a proposal. Rewrite the sentence as a proposed check and remove the attendance claim.” If it omits the exercise, request only that missing component. If it keeps adding unsupported facts after a correction, write the factual paragraph yourself and use AI only to suggest alternative wording. Repeated generation does not make the claim better supported.

Check the claim, the source and the inference

For each factual sentence, underline the words that can be checked: the actor, action, date, number and result. Open the original material and locate the passage that supports those words. A source title or a plausible URL is not enough.

• The source exists but does not support the sentence: narrow the sentence to what it actually says or find a better source. Do not keep an irrelevant citation as decoration.
• The source supports an observation, but the draft asserts a cause: report the observation and explain the uncertainty. In the volunteer example, a wrong meeting point does not prove that it caused lateness.
• Two sources disagree: check their dates, population, method and definition before choosing. Preserve a material disagreement when you cannot resolve it.
• The model supplied a quotation: match it against the original word for word. If no original can be found, remove the quotation; changing its punctuation does not authenticate it.

Keep a short unresolved-claims list next to the draft. A useful entry says “Chapter 3, paragraph 2: confirm whether the map was ever tested; otherwise retain proposal wording.” “Needs fact-checking” is too broad for someone else to finish the job.

Carry the author’s decisions into the next chapter

A voice brief helps most when it records choices that affect meaning. For the volunteer handbook: use direct instructions, avoid treating volunteers as interchangeable staff, distinguish a proposal from an observation, and give a fallback when the coordinator has no extra budget or equipment.

Compare “Create a seamless onboarding experience” with “Before the session, ask someone who has not visited before to find the meeting room using only your welcome email.” The second sentence gives the reader something to try. It also exposes a constraint: if nobody is available to test, the writer needs to provide a weaker fallback and explain what it cannot verify.

Save a chapter handoff, not the whole conversation

Approved facts: N1–N3 remain observations and a proposal; no result has been measured.
Terms: "coordinator" runs the session; "newcomer" is attending for the first time.
Chapter 2 taught: test arrival directions and sign-in instructions.
Chapter 3 must teach: assign a first task after sign-in; do not explain arrival again.
Open issue: confirm who answers questions if the coordinator is unavailable.
Do not invent: volunteer quotations, attendance changes or emotional reactions.

Read the next chapter against this handoff. If it re-teaches arrival, remove the recap unless a reader needs a brief reminder to complete the new task. If it quietly assumes that the map worked, repair that assumption before it spreads into the conclusion.

AI disclosure and publishing requirements

There is no single disclosure sentence that fits every publisher, marketplace, client contract, or jurisdiction. Keep enough production history to answer what material was generated, what was merely assisted, what a human verified, and which rights you hold. Review current requirements close to release rather than relying on an old checklist.

As checked on September 7, 2026, Amazon KDP requires disclosure to KDP of AI-generated text, images, and translations, including generated covers and interior artwork. Substantially editing AI-created content does not change that classification. If you created the content and used AI only to brainstorm, refine, or check it, KDP describes that as AI-assisted and does not require that disclosure. Keep a record of which process you used; these are KDP rules, not a universal rule for every publisher.

• Amazon KDP Content Guidelines: current platform guidance, including its AI-generated and AI-assisted distinction.
• Google Search Central: guidance about generative AI content: emphasizes accuracy, quality, relevance, and people-first value for web publishing.
• NIST AI 600-1: Generative AI Profile: risk-management context for generative systems, including confidently stated false content.
• U.S. Copyright Office: Copyright and Artificial Intelligence, Part 2: a United States-specific report on copyrightability; not a substitute for advice about a particular work.

Frequently asked questions about AI book writing

Test one chapter before drafting the whole book

Select the chapter with the most representative mix of evidence, explanation, example, and reader action. Take it through the entire method and use the defects you find to improve the source packet and section brief. Then follow the separate AI book editing workflow, compare delivery options in the ebook formats guide, or review the broader AI book writer workflow.