“Friendly but professional” is not a voice guide — it’s a wish. Brand voice docs fail because they describe taste without giving anyone (human or AI) a way to apply it. This walks through a workflow that distills voice from real examples into a doc with specific phrases, sentence-length distribution, and “never write this” anti-examples: the kind of doc that actually changes what gets shipped.
TL;DR
- Feed 5-10 of your best on-brand pieces to Claude Sonnet 4.6 or GPT-5.5 and ask for a structured voice breakdown (adjectives, signature phrases, anti-phrases, sentence-length ratio).
- Correct the analysis where it’s wrong, run 3-4 test generations, and tighten the doc each time it drifts. Three rounds usually stabilizes it.
- Keep the final doc to 1-2 pages so it fits inside a Custom GPT (≈8,000-character instruction field, as of June 2026) or a Claude Project (≈8,000-character custom instructions plus up to 200K tokens of uploaded knowledge).
- Re-run the analysis quarterly on your last 90 days of writing. Voice drifts; the doc should too.
What this covers
A reproducible workflow that turns 5-10 of your best pieces of content into a 1-2 page voice doc, then validates the doc by generating test pieces and comparing. The output is loadable into a Custom GPT or Claude Project so every future draft (human or AI) starts from the same voice baseline.
Who this is for
Solo creators producing content under one personal brand. Small marketing teams (2-5 people) where consistency drift is starting to show. Content engineers integrating AI into a content pipeline who need a stable voice prompt to load. Founders who can write in their own voice but can’t yet articulate it to a teammate or freelancer.
When to reach for it
When 2+ people start producing content. When AI starts producing content (the doc becomes the system prompt). When you re-read your own articles from 6 months ago and they sound like a different person. Once the voice doc exists, feed it into a real Claude writing workflow so every draft inherits the same voice. To pressure-test whether a draft actually lives up to that voice doc, run it through AI writing feedback — a senior-editor critique catches the places AI quietly slipped back into AI-grey.
Pick a model and a home for the doc
Two decisions up front, because they shape the format of the doc.
Which model analyzes the samples. Use Claude Sonnet 4.6 or GPT-5.5. Both read 1M-token context windows as of June 2026, so you can paste a dozen long articles into one prompt without truncation. Smaller or older models over-generalize voice into bland adjectives. Claude Opus 4.7 is sharper on subtle voice and humor signal if you have Max access, but Sonnet 4.6 is the cost-effective workhorse here.
Where the finished doc lives. This determines how much room you have:
| Home for the voice doc | Instruction space (June 2026) | Extra knowledge |
|---|---|---|
| ChatGPT Custom GPT (Plus/Pro) | ≈8,000-char instruction field | Upload sample files to the GPT’s Knowledge |
| Claude Project (Pro/Max) | ≈8,000-char custom instructions (~2,000 words) | Up to 200K tokens / 30MB per file, RAG auto-scales on paid plans |
| ChatGPT custom instructions (personal) | ~1,500 chars per field | None — keep the doc tight |
voice.md in a repo / Cursor | No hard limit | Loaded per conversation |
A 1-2 page voice doc (roughly 600-900 words) fits comfortably in all four. That is the real reason to keep it short — not aesthetics, but the instruction-field ceiling.
A note on built-in tone controls: ChatGPT’s 2026 personality presets (Default, Friendly, Efficient, Professional, Candid, Quirky) and the warmth/enthusiasm sliders are blunt instruments. They nudge register; they cannot reproduce your signature phrases or the structures you favor. A distilled voice doc loaded as instructions does what the presets can’t.
Before you start
- Have 5-10 pieces of your best on-brand content. Long-form preferred (blog posts, newsletters) — short copy doesn’t carry enough voice signal.
- Pick ones you genuinely think of as “right.” Mediocre samples will produce a mediocre voice. Brutal selectivity.
- Decide who the doc is for: humans, AI, or both. The format differs slightly — AI prompts need explicit do/don’t lists; human docs can be more narrative.
Step by step
- Collect 5-10 on-brand pieces. Paste them into one document, separated by clear delimiters.
- Run the analysis prompt:
Analyze these pieces of writing for voice. Output:
- 3 adjectives that describe the voice + 3 opposites it avoids
- 5 specific phrases the writer uses (signature words/phrases)
- 5 phrases or words the writer would NEVER use
- Sentence-length distribution: short/medium/long ratio
- Formality level (academic / professional / conversational / casual)
- Joke or humor style if any, with one example
- Common openings (how do paragraphs start)
- Pet structures (parallelism, lists, em-dashes, rhetorical questions)- Review the AI’s analysis. Where it’s wrong, correct it explicitly: “No, we use em-dashes more than this suggests.” The corrections matter more than the initial output.
- Generate test pieces. Prompt: “Using the voice doc above, write a 200-word piece about [a topic you know well].” Read it. Note where it drifts.
- Iterate: when AI drifts, add a rule to the doc that prevents that drift. After 3-4 test generations, the doc stabilizes.
- Save the final doc as a Custom GPT system prompt, a Claude Project context, or a
voice.mdyour team loads into every draft conversation. - Quarterly: re-run analysis on your last 90 days of content. Voice drifts; the doc should too.
Worked example: what a tight doc looks like
Run the analysis on a real archive and the output should read like an instruction sheet, not a personality quiz. A finished entry for a developer-tools newsletter, after two correction rounds, looks like this:
VOICE: direct, dry, technically precise. AVOIDS: hype, salesy, vague.
SIGNATURE PHRASES: "here's the catch", "in practice", "the boring answer is",
"ship it", "this breaks when".
NEVER WRITE: "game-changer", "revolutionize", "unlock", "in today's
fast-paced world", "leverage" (as a verb).
SENTENCES: ~60% short (<12 words), 30% medium, 10% long. One-sentence
paragraphs are fine for emphasis.
FORMALITY: conversational-technical. Second person ("you"), contractions yes.
HUMOR: deadpan understatement, never puns. e.g. "The migration took a
weekend. The bug took three."
OPENINGS: start with the problem or a claim, never a definition.
STRUCTURES: em-dashes for asides (sparingly), numbered steps, no rhetorical
questions.The test is whether someone who knows the writing reads that block and says “yes, that’s us.” If the signature phrases are generic (“we value quality”), the doc is still describing a wish.
Quality check
- Are the “5 signature phrases” actually phrases from your writing, or generic ones the AI assumed? If you can grep your archive and not find them, the AI hallucinated.
- Are the “never use” rules specific enough to be checkable? “Don’t be too formal” is unverifiable; “never use ‘leverage’, ‘utilize’, or ‘synergy’” is checkable.
- Does the test generation pass a 5-second sniff test from someone who knows your writing? If they say “this isn’t you,” the doc isn’t tight enough yet.
- Is the doc short enough to fit the instruction field and still get read? Aim for 1-2 pages (600-900 words). A 10-page voice doc is ignored, and it won’t fit the ≈8,000-character ceiling anyway.
Reuse it across tools
- Save the analysis prompt as a template. Re-run it quarterly on new samples; voice evolves.
- Maintain a “rejected outputs” log: AI generations that drifted, with notes on why. These become future negative examples — the anti-phrase list grows with use.
- Load the same doc into every AI writing tool you touch — Custom GPT instructions, a Claude Project, and a
voice.mdyou paste into Cursor for marketing copy. One source of truth, several integrations. - Before the voice doc is fully useful, the team also needs one shared positioning anchor. Pair this tutorial with our AI positioning statement workflow so voice and message stop drifting against each other.
Common mistakes
- Defining voice from 2 examples — too thin. The signal is in the patterns across samples, not any single piece.
- No regular update — voice drifts and the doc becomes archaeology. Quarterly refresh.
- Voice doc that’s 10 pages — no one reads it, the AI loads an unreadable instruction block, and it overflows the ≈8,000-character field. Edit ruthlessly.
- Using only “great” pieces without bad examples — anti-examples are half the signal. Include 1-2 pieces the AI should NOT sound like.
- Trusting the AI’s “joke style” analysis without verification — humor is the hardest signal; AI often misreads it.
- Stopping at the voice doc when the creator is a person, not a brand — extend with AI creator brand tutorial so the doc captures founder voice, not generic brand voice.
FAQ
- Which model should run the analysis?: Claude Sonnet 4.6 or GPT-5.5 (both 1M-token context as of June 2026). Claude Opus 4.7 reads subtle voice and humor best if you have Max. Avoid smaller models — they flatten voice into generic adjectives.
- How big can the voice doc be before it stops fitting?: A Custom GPT instruction field and a Claude Project’s custom instructions both hold roughly 8,000 characters (~2,000 words) as of June 2026. ChatGPT’s personal custom-instruction fields are tighter at ~1,500 characters each. Keep the doc to 600-900 words and it fits everywhere.
- Can’t I just use ChatGPT’s tone presets instead?: The 2026 presets (Friendly, Professional, Candid, Quirky) and warmth sliders adjust register, not your specific phrasing. They don’t reproduce signature phrases or structures. Use a loaded voice doc for that.
- Can the voice doc replace a copy editor?: No — not for facts, structure, or first drafts. It enforces tone consistency, which is one slice of editing.
- My voice changes by channel (newsletter vs. social vs. landing pages). One doc?: Multiple docs, one per channel, cross-linked in a master index. A landing page and a personal newsletter rarely share sentence-length ratios.
- How do I share this across a team?: Voice doc in the team handbook, loaded as Claude Project context, plus a 30-minute walkthrough on the rationale. The walkthrough matters more than the doc — people apply rules they understand.
For the platform limits referenced above, see Anthropic’s Projects overview and OpenAI’s custom instructions guide.