Galley
FIELD GUIDE NO. 04

Writing a Book with AI Without Losing Your Voice

Use AI to accelerate the work while preserving human judgment, evidence, originality, and authorship.

An eight-step loop · An eight-part scorecard · A 90-minute sprint

There are two ways to lose a book to AI. The obvious one is letting it write badly. The quiet one is letting it decide.

The first failure is visible: generic openings, invented statistics, a voice that belongs to nobody. The second is worse because the prose looks fine. The model chose what the chapter argues, which objections got answered, and which examples appeared, and the author became an editor of text they never actually thought. The drafting got faster, and the authorship leaked away one accepted suggestion at a time.

Using AI well in nonfiction is not a matter of better generation prompts. It is a boundary: the author owns the promise, the argument, the evidence, and the final expression, and the model works inside clearly defined tasks with verification behind every one. This guide is that boundary made operational: the authorship modes, the voice charter, the eight-step draft loop, and the records that let you answer KDP's disclosure questions with confidence instead of guesses.


01 THE SHORT ANSWER

What AI should and should not do

The safest and most useful role for AI in nonfiction is not write my book. It is help me complete a clearly defined part of the work while I remain responsible for the reader, argument, evidence, language, and final text. AI is useful when the author has already defined the problem; its apparent fluency becomes dangerous when fluency is mistaken for truth, originality, or judgment.

STRONG USES

Generating alternative explanations or structures. Clustering research notes and reader questions. Identifying missing prerequisites or objections. Turning an approved outline into section briefs. Producing counterarguments to evaluate. Diagnosing repetition, vagueness, and weak transitions. Proofreading and flagging inconsistencies.

WEAK OR DANGEROUS USES

Deciding what the book should promise without market evidence. Inventing facts, sources, quotations, statistics, or experiences. Summarizing material the author has not read. Imitating a recognizable living author. Generating a whole manuscript from a broad prompt. Declaring text original or legally safe. Replacing expert review in a consequential subject.

THE ACCOUNTABILITY TEST

If this output is wrong, misleading, copied, biased, private, or off-promise, do I know how I will detect it?

If the answer is no, narrow the task, add evidence, or keep the work human.

A human-led process has six non-negotiables: the reader promise and outline exist before generation begins; every section has a human-authored purpose and evidence plan; AI output is treated as unverified material, not finished prose; the author's examples, decisions, language, and point of view shape the final text; claims, quotations, and citations are verified against original sources; and AI use is recorded accurately enough to answer KDP's disclosure questions. The governing principle: never delegate a decision you cannot later explain, verify, and defend.


02 DECIDE FIRST

Choose the authorship mode before drafting

Do not wait until the KDP upload screen to reconstruct how the manuscript was made. Decide the workflow in advance and keep a simple record.

MODE A · HUMAN-WRITTEN, AI-ASSISTED

The author creates the actual prose. AI helps brainstorm, outline, edit, refine, or error-check it. Under KDP's current definitions this is AI-assisted, and no disclosure is required.

MODE B · AI-GENERATED, HUMAN-DIRECTED

An AI system creates actual text used in the book. The author prompts, selects, fact-checks, and edits it. KDP treats this as AI-generated, even after substantial edits, and it must be disclosed.

MODE C · MIXED MANUSCRIPT

Some passages are human-written and AI-assisted; others begin as AI-generated prose that stays in the book. Track the relevant text as AI-generated.

Keep an authorship ledger with one row per section: the starting mode, the AI task, the human contribution, and the KDP classification. The ledger is not a public confession or a measure of literary worth. It is operational memory, and it prevents guesses when the platform asks how text, images, or translations were created.


03 VOICE

Build the Author Voice Charter

Make it sound human is not a voice strategy. Define the choices that make the book sound like this author speaking to this reader for this purpose. Start from 1,000 to 3,000 words of writing you own and have the right to process, with confidential material removed, and describe eight dimensions.

01

Reader relationship

A guide beside the reader, an instructor at the board, a practitioner reporting from the field, or a challenger confronting assumptions.

02

Expertise stance

Whether you speak from research, direct experience, or practice, and where the boundary lies between what you know and what remains uncertain.

03

Sentence rhythm

Mostly short direct sentences, longer analytical ones, or deliberate variation, and the typical paragraph length.

04

Vocabulary

Words natural to you and familiar to the reader, and the jargon, cliches, and inflated verbs to avoid.

05

Explanation pattern

Story to principle, problem to method, claim to evidence, or instruction to example, chosen as a repeatable rhythm.

06

Evidence stance

How you qualify uncertainty, distinguish observation from general claim, and cite changing information.

07

Emotional range

How warm, urgent, skeptical, humorous, or encouraging the voice is, with limits so tone does not swing between sections.

08

Signature moves

Techniques that recur naturally: diagnostic questions, compact analogies, numbered processes, field notes, or end-of-section actions.

INSTEAD OF IMITATION

Write this like a famous bestselling habits author.

DESCRIBE ATTRIBUTES YOU OWN

Use short paragraphs, concrete examples, plain verbs, one central idea per section, and a practical action at the end. Avoid slogans, invented stories, and exaggerated certainty.


04 EVIDENCE

Prepare the evidence packet

Do not ask a model to research and write in one step. Separate evidence collection from prose generation, and give every section a packet before drafting begins.

  • The approved chapter card and section job.
  • Verified facts and claims.
  • Primary or authoritative source links.
  • Exact quotations with location and permissions status.
  • The author's observations and experiences.
  • Interview notes with consent and attribution needs.
  • Examples or case material you may use.
  • Known disagreements or limitations.
  • Facts that may change, with a checked date.
  • Material that must not be included.

Inside the packet, keep four kinds of statement distinct. A verified fact has a source that directly supports it. An author observation is what you saw, did, or measured, with its scope signaled: in my projects is not research proves. An interpretation is a conclusion drawn from evidence, with the reasoning shown. A recommendation is advice to act, with conditions, trade-offs, and limits. Track them in a claim ledger with the preferred evidence, whether the source was opened, and a status of verified, qualify, or remove. AI may help organize these categories. It cannot promote a claim from possible to verified.


05 THE METHOD

The eight-step Human-First Draft Loop

Draft one section at a time. A section is small enough to direct, verify, rewrite, and integrate without losing the book's logic.

1

Own the section job

Complete this sentence before prompting: by the end of this section, the reader will understand, decide, create, perform, diagnose, or avoid one specific thing. If you cannot write it, return to the outline. AI-generated prose cannot repair an undefined purpose.

2

Write the human brief

Specify the reader's entry condition, the one job, required claims and sources, the author's point of view, the example or experience, the reader action, the boundary of the section, and the transition to what follows. This brief is the editorial contract the model is not allowed to silently change.

3

Create the human source notes

Before generation, write rough notes in your own words. These preserve authorship at the level of ideas and judgment, not only surface wording.

  • What do I believe this section must say, and what have I personally observed?
  • What is counterintuitive or commonly misunderstood?
  • What evidence changes or limits the advice, and what example can only I provide?
4

Ask AI for options, not authority

Propose three explanation sequences, identify a missing objection, turn verified notes into a table, critique the logic, or draft a provisional version using only the supplied evidence. Label the result unverified working material.

5

Draft or reconstruct in your own voice

Do not limit revision to synonyms. Rebuild the section around your intent: choose the opening, decide what deserves emphasis, remove generic throat-clearing, add the example and trade-off, and use your natural vocabulary and rhythm. If AI-generated prose remains in the book, record it as AI-generated even after a substantial rewrite.

6

Add earned specificity

Generic prose sounds reasonable because it avoids conditions. Add details that make advice testable: who should do it, when it applies, what inputs are needed, what done looks like, what can go wrong, and what the method does not solve. Specificity should come from evidence and experience, not invented precision.

7

Verify every consequential element

Check facts and numbers; names, dates, places, and quotations; citations and links; causal claims; high-stakes advice; descriptions of people, products, and studies; permissions and privacy; and whether examples are real, composite, or hypothetical. Open the source. Locate the exact support. Do not treat a plausible citation as evidence.

8

Integrate and record

Read the section in context, then update the master manuscript, the claim ledger, the authorship ledger, the source list, the consistency sheet, and the outline if the section's job changed. The loop is complete only when the section belongs to the book, not merely when it contains polished sentences.


06 IN PRACTICE

A worked drafting example

The remote-management book from the earlier guides reaches drafting. The section job: a first-time remote manager can define a response-time agreement that protects focus without leaving urgent work unattended.

WEAK ONE-LINE PROMPT

Write 800 words about communication for remote teams in a friendly style. The result: a communication is key opening, generic tool lists, invented statistics, no artifact the reader can implement, and no connection to the surrounding chapter.

BETTER AI TASK

Using only the supplied notes, identify three possible teaching sequences for explaining response-time agreements. For each, show the reader question answered, the manager decision required, and the practical output. Do not write the section, invent evidence, recommend software, or claim a universal response time.

GENERIC VERSION

Effective communication is essential for every successful remote team. Managers should set clear expectations, use the right tools, and encourage open communication so everyone stays aligned.

AUTHORED VERSION

A green status light answers: are they online? It does not answer: when should they reply? When a team never separates those questions, every message can feel urgent, even when nobody intended it that way. A useful response-time agreement names three things together: the message class, the expected window, and the escalation path if waiting would cause harm.

The stronger version is not better because it sounds less like AI. It is better because it makes a distinction, explains a mechanism, and prepares the reader to build something specific.


07 INTEGRITY

Fact-checking, citations, and source integrity

Generative AI can produce confident falsehoods, fabricated citations, and correct-looking details in the wrong context. NIST uses the term confabulation for confidently presented false or erroneous content, including fabricated logic and citations. Four rules keep the manuscript honest.

THE SOURCE-OPEN RULE

Never cite a source you have not opened. Confirm it exists, confirm author, publisher, date, and title, locate the exact passage, check it supports the sentence as written, read enough context to catch qualifications, prefer the primary source, and record the access date.

THE QUOTE RULE

Copy every quotation from the original source, not an AI output. Preserve wording and context, identify the speaker accurately, record the location, determine whether permission is needed, and keep quoted material proportionate to the purpose.

THE HIGH-STAKES RULE

Medical, legal, financial, safety, and psychological advice deserves domain-expert review. A disclaimer cannot repair reckless content, and an AI system is not an accountable substitute for a qualified professional.

THE FRESHNESS RULE

Mark facts likely to change: laws, platform policies and interfaces, prices, product capabilities, statistics. Include a checked date and create an update schedule before publication.


08 RIGHTS

Originality, copyright, and responsible reuse

This is practical publishing guidance, not legal advice, and it starts from an uncomfortable truth: rewriting words is not automatic permission. Facts and ideas may be available to discuss, but protected expression, distinctive examples, structure, and images can create rights issues, and replacing words with synonyms does not make a use lawful, ethical, or original. Research several sources, develop your own analysis, cite appropriately, and obtain permission where required.

NO UNIVERSAL SAFE PERCENTAGE

Do not use a plagiarism-checker score as legal clearance. The US Copyright Office states that fair use has no formula based on a predetermined percentage or number of words, and only a court can finally determine it. Similarity tools flag passages for review; they cannot decide infringement.

DETECTORS ARE NOT AUTHORSHIP RECORDS

Do not rely on an AI detector to prove text is human. Preserve the real workflow: dated notes, version history, source packets, prompt records where appropriate, tracked revisions, the ledgers, and editorial review records.

HUMAN AUTHORSHIP AND US COPYRIGHT

The US Copyright Office's current position is that material generated wholly by AI is not copyrightable. AI-assisted human creation can still qualify where a human contributes sufficient expressive elements through their own material, arrangement, or modification. Prompts alone are not enough.

And before uploading anything to an AI system, confirm five things: you own it or have permission to process it, it contains no confidential or restricted information, personal data is minimized or removed, client, interviewee, and employer obligations are respected, and the provider's current terms, data controls, and training settings fit the use.


09 DECISION AID

The human-authorship scorecard

Score each dimension from 0 to 2, and add one sentence of evidence beside each score.

Human ownership

0 · WEAK

AI determines purpose, reasoning, and prose

1 · MIXED

Human directs some choices but accepts major generated decisions

2 · STRONG

Human owns the promise, argument, structure, examples, and final expression

Voice fidelity

0 · WEAK

Generic or inconsistent voice

1 · MIXED

Voice appears in places but drifts

2 · STRONG

The Voice Charter is evident and consistent

Reader usefulness

0 · WEAK

Fluent information without a clear job

1 · MIXED

Some practical value, with vague sections

2 · STRONG

Every section produces understanding, a decision, action, or artifact

Earned specificity

0 · WEAK

Generic claims and invented precision

1 · MIXED

Some real examples and conditions

2 · STRONG

Evidence, experience, boundaries, and trade-offs make advice specific

Evidence integrity

0 · WEAK

Claims or citations are unverified

1 · MIXED

Most important claims checked

2 · STRONG

Consequential claims map to opened, authoritative sources

Disclosure traceability

0 · WEAK

AI use cannot be reconstructed

1 · MIXED

Major AI use is remembered but incompletely logged

2 · STRONG

Authorship mode is recorded section by section

Rights and privacy

0 · WEAK

Source, permission, or confidentiality risks remain

1 · MIXED

Screening exists with unresolved items

2 · STRONG

Rights, reuse, personal data, and permissions are controlled

Book integration

0 · WEAK

Sections feel independently generated

1 · MIXED

Local coherence with some repetition

2 · STRONG

Terms, examples, sequence, and transitions serve one unified book

RUNNING TOTAL
/16

13 to 16 means AUTHOR: proceed to editorial and expert review. 9 to 12 means REWORK: preserve the section job but repair weak authorship, evidence, voice, or integration. 0 to 8 means RECLAIM: stop generation and rebuild from the human brief, source notes, and Voice Charter. An unresolved rights, privacy, disclosure, safety, or serious factual issue overrides the score.

THE PRINTED EDITION

Take the draft loop with you.

This page has the method. The designed field guide has it in the form you actually use: twenty-seven pages built to be printed, with the Voice Charter and section brief as working forms, the full five-prompt stack, the claim and authorship ledgers, and the disclosure gate. Same content, better working copy.

The download opens straight away and a copy arrives by email. We send a note when a new field guide lands, and you can leave any time.

10 WATCH OUT

Common AI writing mistakes

Asking for the whole book in one prompt

The output may be coherent sentence by sentence while drifting across chapters, repeating advice, and inventing bridges.

Beginning without a validated outline

AI can fill empty structure so quickly that the author mistakes volume for progress.

Prompting in the style of a named living author

Translate admired craft into explicit attributes you can own instead of requesting imitation.

Letting AI research and cite in one pass

This invites fabricated or misapplied sources. Separate discovery, source opening, drafting, and citation verification.

Editing only for human-sounding wording

Replacing cliches or varying sentence length does not add judgment, evidence, lived specificity, or a point of view.

Treating confident language as expertise

Fluency can hide uncertainty. Ask what supports the claim and what would disprove it.

Assuming heavy editing removes the disclosure obligation

Under KDP's current definition, AI-generated text remains AI-generated even after substantial edits.

Inventing personal experience through AI

Never allow a model to fabricate what you saw, felt, tested, achieved, or learned.


11 FAST START

The 90-minute section sprint

Use this for one substantial section after the outline, Voice Charter, and evidence packet are ready.

0 TO 10 MIN · DEFINE THE JOB

Complete the section brief. Confirm the reader's entry and exit conditions, and state the boundary.

10 TO 25 MIN · ASSEMBLE NOTES AND EVIDENCE

Write your point of view in rough language. Add experience and examples. Open the sources and update the claim ledger.

25 TO 35 MIN · USE AI FOR OPTIONS OR CRITIQUE

Generate teaching sequences, objections, or a provisional draft per the chosen mode. Mark all output unverified.

35 TO 60 MIN · AUTHOR THE SECTION

Choose the reasoning and order. Write or reconstruct in your voice. Add earned specificity, limits, and the reader action.

60 TO 75 MIN · VERIFY

Extract claims and open sources. Check quotations, names, numbers, and causal statements. Remove or qualify unsupported material.

75 TO 90 MIN · INTEGRATE AND RECORD

Read the surrounding sections and repair repetition and transitions. Update the authorship, claim, and source records. Save the version and the next action.

Finishing the timer with unresolved evidence does not make the section complete. Mark it clearly and return after verification.


12 CARRY IT FORWARD

The KDP disclosure and quality gate

Run this gate before the manuscript advances to formatting. The disclosure decision reduces to two questions.

DISCLOSE · AI-GENERATED

Did an AI tool create actual text, images, or translations used in the published book? Then the content is AI-generated and must be disclosed during publishing, even after substantial edits.

NO DISCLOSURE · AI-ASSISTED

Did you create the content yourself while AI only brainstormed, edited, refined, or error-checked? Then it is AI-assisted under Amazon's current definition, and no disclosure is required.

If the answer is uncertain or mixed, review the authorship ledger, preserve the more cautious classification, and consult current KDP guidance. Disclosure does not remove the publisher's responsibility for accuracy, rights, and customer experience. Then close the quality gate.

  • Every section fulfills an approved job in the book's blueprint.
  • The Voice Charter is consistent across chapters.
  • Personal experiences are real, authorized, and accurately represented.
  • Hypothetical, composite, and anonymized examples are labeled appropriately.
  • Facts, quotations, citations, names, and numbers were checked against original sources.
  • High-stakes advice received qualified expert review where needed.
  • Changing facts include a checked date and a maintenance plan.
  • Unsupported claims were removed or qualified.
  • Rights, permissions, confidentiality, and personal data were reviewed.
  • Repeated passages, recycled examples, and chapter contradictions were removed.
  • The authorship ledger supports the KDP disclosure answer.
  • The entire manuscript, not only sampled sections, received human editorial review.

The reader did not buy access to sentences. They bought a trustworthy route through a problem.

The section sprint takes 90 minutes, and most of it is not generation. It is briefing, verifying, and authoring, which is the point: the model accelerates the sentences, and the sentences were never the hard part. The hard part is knowing which book deserves the effort and what each section owes the reader.

Galley runs the page-one checks and returns a verdict in about a minute, then carries the validated promise into a Book Blueprint of section briefs, evidence packets, voice controls, and authorship records. AI can assist inside that system without becoming the system.

Free. No card.

13 REFERENCE

Frequently asked questions

Does Amazon KDP allow AI-generated books?
KDP currently permits AI-generated and AI-assisted content that follows its content guidelines. Publishers must disclose AI-generated text, images, or translations during publishing; they do not need to disclose AI-assisted work under Amazon's current definition. The publisher remains responsible for rights, accuracy, and customer experience.
What counts as AI-generated versus AI-assisted on KDP?
AI-generated content is actual text, imagery, or translation created by an AI tool and used in the book, and it remains AI-generated even after substantial editing. AI-assisted content is created by the author while AI helps brainstorm, edit, refine, or error-check.
If I heavily rewrite AI text, do I still have to disclose it?
Under KDP's current definition, yes. If the tool created the actual content, Amazon considers it AI-generated even when the publisher applies substantial edits. Track it in an authorship ledger so the answer is a record, not a memory.
Can AI write my entire book?
An AI system can produce manuscript-length text, but length is not authorship, accuracy, market fit, coherence, or quality. Whole-book generation multiplies the burden of verification, integration, rights review, disclosure, and human contribution. A section-by-section human-led workflow is more controllable.
How do I stop AI writing from sounding generic?
Do not solve the problem with tone words alone. Supply a validated section job, human source notes, a voice charter, evidence, examples, boundaries, and a reader action, then make the decisions and final expression yourself. Generic prose is usually a missing-inputs problem, not a phrasing problem.
Should I ask AI to write like a famous author?
Avoid requesting imitation of a named living author. Describe the craft attributes you need, such as sentence rhythm, vocabulary, evidence stance, explanation pattern, and tone boundaries, then develop a voice grounded in your own writing.
Can I trust citations generated by AI?
No citation should be trusted until you open the original source, confirm it exists, locate the exact support, and check the surrounding context. Fabricated and misapplied citations are a known generative-AI risk.
Is AI-generated writing protected by copyright?
In the United States, the Copyright Office currently states that wholly AI-generated material is not copyrightable. AI-assisted work can still qualify where sufficient human-authored expression, arrangement, or modification exists, and prompts alone are not enough. Obtain legal advice for a specific manuscript or registration decision.

Your voice is not a tone setting. It is the pattern of choices only you are accountable for making.

Direct the tool. Verify the work. Author the book.

SOURCES AND OFFICIAL REFERENCES
  • Amazon KDP Content Guidelines

    Current definitions and disclosure requirements for AI-generated and AI-assisted text, images, and translations.

  • Amazon KDP: Guide to Kindle Content Quality

    Expectations concerning misleading, duplicated, incorrect, and poor-quality content.

  • US Copyright Office: Copyright and Artificial Intelligence, Part 2

    Current US analysis of human authorship, AI-generated material, prompts, arrangement, and modification.

  • US Copyright Office: Fair Use FAQ

    Why no fixed word count or percentage provides automatic fair-use clearance.

  • NIST: Generative AI Profile

    Risk-management guidance, including generative-AI confabulation and fabricated citations.

Platform policies, AI systems, provider terms, and legal standards can change. Recheck official guidance before publication, and obtain qualified professional advice for consequential rights, privacy, safety, or compliance questions.