Brand Voice Prompt

What this prompt does

A brand voice prompt turns three to five samples of your best copy into a written voice guide — tone positions with quoted evidence, vocabulary rules, sentence mechanics, and a condensed block you can paste at the top of any other prompt. Every piece of advice about keeping AI output on-brand ends the same way: write a short voice document and reuse it everywhere. The advice is right. It just skips the hard part, which is that describing your own voice from a blank page is genuinely difficult: you're too close to it, and everything sounds "friendly but professional" from the inside.

So this prompt doesn't ask you to describe anything. It asks for evidence — real pieces of copy that already sound like you — and derives the description, quoting its receipts as it goes.

It's a different job from fixing tone. The tone adjustment prompt rewrites one message whose register came out wrong. This one produces the reference document that stops the register coming out wrong in the first place. You run it once per brand, then spend the output for months.

The prompt

Reverse-engineer a brand voice guide from real copy. Work only from evidence in the samples below.

**What the brand sells, and to whom:** [ONE OR TWO SENTENCES]
**Samples that sound like us at our best:** [PASTE 3–5 PIECES — an email, a product page, a caption. Variety beats volume.]
**Samples that sound wrong for us:** [OPTIONAL — PASTE 1–2, AND SAY WHAT FELT OFF IF YOU CAN]
**Non-negotiables:** [BANNED WORDS, REQUIRED LEGAL PHRASING — or "none"]

Build the guide in five parts:

1. Voice summary — three sentences maximum. No lists of adjectives.
2. Tone position on four scales: funny–serious, formal–casual, irreverent–respectful, enthusiastic–matter-of-fact. For each scale, state where this writing sits and quote the line from my samples that proves it.
3. Vocabulary — words and constructions the samples actually use, and words this brand would never use. Anything inferred rather than observed gets marked [inferred — confirm].
4. Sentence mechanics — typical sentence length, contractions or not, first or second person, punctuation habits, how pieces open and how they close.
5. A paste-ready block — 120 words or fewer, condensed from parts 1–4 and written as instructions ("Write in..."), ready to sit at the top of any prompt.

Rules:
- Every claim in parts 2–4 quotes its evidence from the samples. No evidence, no claim — or mark it inferred.
- Generic descriptors are banned: "friendly," "professional," "authentic," and "engaging" describe every brand and therefore none. Say what the writing does instead.
- Don't flatter the samples, and don't invent brand values that aren't on the page. If two samples contradict each other, say exactly where.

How to use it

Sample selection decides everything, because the model can only see what you show it. Pick pieces you'd happily write again, from more than one format — an email plus a product page plus a social caption reveals more mechanics than three emails. If a piece got a great response but doesn't sound like you, leave it out; you're documenting the voice, not the greatest hits.

The wrong-samples input earns its line. One example of copy that felt off — too hypey, too stiff, a freelancer draft you heavily rewrote — gives the model a contrast to define the boundary against, and the "would never use" list comes back sharper for it.

Then edit the output like the draft it is. The guide can only describe who you've been in those samples, not who you're trying to become — if the summary reads accurate but you wince at it, that's a positioning decision the prompt correctly refused to make for you.

What lands in the guide, and where it comes from

The four tone scales aren't arbitrary. Nielsen Norman Group's four dimensions of tone of voice — humor, formality, respectfulness, enthusiasm — exist precisely because adjective lists fail: every brand calls itself friendly and professional, so the words carry no information. A position on a scale, tied to a quoted line, does.

Example output (partial)

Inputs: a specialty coffee subscription, three samples pasted — a shipping-delay email, a bag label, an Instagram caption.

Tone positions. Serious-leaning with dry asides, not funny: "Your beans are late. We're not going to pretend that's fine." Casual: contractions in all three samples, sentences average 9 words, one-sentence paragraphs used for emphasis. Respectful, no irreverence toward the customer — the joke is always on us ("we overslept the roast"). Matter-of-fact over enthusiastic: zero exclamation marks across 340 words of samples.

Would never use: "artisanal," "curated," "elevate," "coffee journey." The bag label describes origin and process in plain nouns — "washed, dried on raised beds" — so [inferred — confirm: the brand avoids all category clichés, not just these four].

Paste-ready block (excerpt): Write in first person plural, second person to the reader. Sentences short, average under 12 words. Contractions always. No exclamation marks. When something goes wrong, state it plainly in the first line and make any joke at our expense, never the reader's...

That zero-exclamation-marks line is the kind of thing you'd never think to write in a voice doc from scratch, and it's exactly the instruction that keeps a model from drifting into cheerleader register on post three.

Variations

For an agency running multiple clients

Run it once per client and keep each guide in its own file. The per-brand setup is a one-time cost of maybe half an hour including your edit pass; after that, switching voices is swapping which block you paste. When a client's copy is thin, borrow their founder's emails as samples — voice usually starts there anyway.

For auditing drift

Reverse the direction: paste the finished guide plus a new draft, and ask for every line that violates the guide, quoted, with the rule it breaks. This catches the slow slide back toward model-default polish that no one notices post by post.

For a brand that has no copy yet

You can still run it — paste positioning notes and two or three samples from brands whose voice you admire, state how you want to differ, and treat the entire output as [inferred — confirm]. What comes back is a hypothesis, not a description. Write ten real pieces with it, then rerun the prompt on the ones you kept.

Common pitfalls

• Don't paste your mission statement as a sample. It's copy about the brand, written in boardroom voice, and it will drag every tone position toward formal-and-enthusiastic — the opposite of most brands' actual writing.
• Don't mix five writers' work in one sample set and expect one voice back. The prompt will tell you the samples contradict each other, which is useful information, but it's an audit finding, not a guide.
• Don't let a generic descriptor survive your edit pass, even one the model smuggled in as a quote. The banned-word rule exists because the model drifts back to "friendly and professional" whenever the evidence is thin.
• Try instead: before trusting the guide, have the model write one new piece using only the paste-ready block, and read it next to a real sample. Where the new piece sounds wrong, the block is missing a rule — add it and test again.

Who uses this prompt

• Copywriters: building a voice reference at the start of a client engagement, so every draft after the first needs less rewriting. More in prompts for copywriters.
• Marketers: keeping campaign copy on-voice across channels and freelancers. See prompts for marketers.
• Social media managers: maintaining distinct voices across client accounts, where the caption prompt does the daily writing and this guide keeps it from flattening into one house style. See prompts for social media managers.

ChatGPT, Claude, and Gemini all run this unmodified, though in my testing Claude quotes evidence more readily while ChatGPT needs the no-evidence-no-claim rule restated once. The mechanism is the one Anthropic's prompt engineering guidance keeps returning to: examples beat descriptions. Which is the quiet point of the whole exercise: the guide works not because the model understands your brand, but because you finally showed it three pieces of proof.