To give AI your brand voice in every chapter, write a one-page voice sheet with measurable traits, attach two or three samples of your real writing, add a short ban list, and paste that same block into every chapter prompt. Then check a few simple numbers to confirm the voice held.
Below you'll find what a voice sheet contains, an example with real measured numbers, a ban list approach based on published research, and a drift check you can do in minutes.
Why doesn't "write in a friendly, professional tone" work?
Because those words mean different things in every chapter. "Friendly" has no length, no rhythm and no vocabulary attached. The model picks an average of what friendly usually looks like, and that average is the generic tone readers recognize as machine-written. Anthropic's guidance on examples notes that examples are "one of the most reliable ways to steer" output format, tone and structure, and Google's guide recommends specific, varied examples as well. Samples and numbers carry more information than adjectives.
This is the same reason that telling the AI to be an expert has limited effect: labels are cheap, details are expensive. Give the model details.
What goes into a voice sheet?
Six to eight observable traits, each checkable by a reader or a script. Here is a template with an example filled in:
| Trait | How to specify it | Example |
|---|---|---|
| Person | Who is speaking to whom | "I" for opinions, "you" for the reader |
| Sentence length | Average and cap | Average about 16 words, none over 35 |
| Contractions | Yes or no | Yes: "you'll", "it's" |
| Openers | How sections start | One-sentence answer in bold |
| Examples | What kind and how many | One concrete example per section, with numbers |
| Humor and emotion | Allowed or not | Dry, rare, never in instructions |
| Formatting habits | Lists, tables, headings | Question-style H2s; tables for comparisons |
| Never | Phrases to avoid | See ban list below |
Where do the numbers come from? Measure your own best writing. As an illustration, I ran a short script on this site's rewritten guide to prompt length (skipping tables and code): it came out at an average of 18.4 words per sentence, a median of 16.5, about 1.4 contractions per 100 words and about 1.8 uses of "you" per 100 words. Those are the voice traits of that article, not a standard to copy. Your numbers will differ, and the point is that you can write them down and re-measure.
How do I choose sample paragraphs?
Pick two or three short passages that you'd happily publish, from different topics. Mixed topics prevent the model from linking the voice to one subject. Keep each at 100-250 words, label them clearly (for example inside <voice_sample> tags), and add a one-line note on what each demonstrates. The structure of the prompt follows the pattern in few-shot prompting for consistent chapters: instruction, examples, anti-copy rule, then the request.
<voice_sheet>
[traits from the table]
</voice_sheet>
<voice_sample note="direct opener, one number, short sentences">
[paragraph]
</voice_sample>
<voice_sample note="story opener, then the point">
[paragraph]
</voice_sample>
Match the voice, not the topic or phrases. Never use: [ban list].
If you have no writing of your own yet, write one chapter yourself or heavily edit a draft, then use it as the first sample. Using someone else's writing as your voice sample raises copyright and originality problems, so don't.
What goes on the ban list?
The words that make text sound generic, plus any of your own pet peeves. Research gives a useful starting point: Kobak et al. analyzed over 15 million PubMed abstracts and found that after ChatGPT arrived, certain style words surged in frequency, with "delves" among the strongest and "underscores" and "showcasing" also elevated. That's evidence about scientific abstracts, not your genre, so test the candidates against your own drafts. A starting list for nonfiction might include: delve, tapestry, landscape (figurative), unlock, leverage, game-changer, "in today's fast-paced world", "it's important to note". Keep it under 20 items; long ban lists crowd the prompt and invite odd substitutions. The full routine is in how to make AI writing sound less like AI.
Worked example: turning a vague voice into a spec
Replace each adjective with a rule you could check. The mapping below is an example I wrote for a fictional personal-finance guide for freelancers:
| Vague brief | Checkable spec |
|---|---|
| "Warm" | Uses "you" and "we"; at least one reassuring sentence per section; contractions allowed |
| "Authoritative" | States the answer first; gives a number or a source for each major claim; no hedging stacks like "might possibly" |
| "Not too corporate" | No "leverage", "synergy", "stakeholders"; prefers "use", "work with", "people" |
| "Conversational" | One short question per section; sentences vary between 5 and 30 words |
| "Practical" | Every section ends with an action that takes under 10 minutes |
Once the right-hand column exists, you can give it to the model, check it yourself, and hand it to an editor. The spec also survives a change of tool or a fresh chat, which a feeling does not. Start the whole project from that sheet: in Interactive eBook Builder, the Seed Prompt initializes the eBook's overall structure and can be re-run to repair the workspace or start the next eBook, so the setup work is repeatable rather than rebuilt from scratch for each book.
How do I check for drift between chapters?
Count the same four numbers for each chapter and compare them to your voice sheet. It takes a few minutes with a spreadsheet or a small script.
| Metric | Target (example) | Chapter 2 | Chapter 8 | Action |
|---|---|---|---|---|
| Average sentence length | 16 +/- 3 | (your count) | (your count) | Outside range: ask for shorter sentences |
| Contractions per 100 words | 1.0-2.0 | (your count) | (your count) | Zero: voice is stiffening |
| "You" per 100 words | 1.5-2.5 | (your count) | (your count) | Low: text turned abstract |
| Ban-list hits | 0 | (your count) | (your count) | Replace and re-prompt |
I've left the chapter columns blank because they depend on your book. Numbers can't judge voice entirely (humor and warmth don't show up in counts), so read a paragraph aloud too. If chapters drift, the usual cause is that the voice block was edited between prompts, so freeze it and version it.
A staged tool can help hold this still. In Interactive eBook Builder, the agent instructions are editable and saved in one place, and Agent C Polisher includes a voice pass, so the same rules are applied to the finish of every chapter instead of being retyped. Whichever tool you use, the principle is identical: one source of truth for voice, reused everywhere. You can also place the same voice block in a reusable instructions page, as covered in system prompts for a Notion AI writing assistant.
Before you hit send
- Is your voice sheet written in measurable traits, not adjectives?
- Are there two or three samples from different topics?
- Did you add a ban list of under 20 items and an anti-copy line?
- Are you pasting the identical block every time?
- Did you compare a few numbers to chapter 1 after drafting?
The takeaway: voice is a spec, not a mood. Write it down in numbers and samples, reuse it unchanged, and measure it.