Brand Voice — Freeze your voice into a reusable file

Take real samples of what you have already published and distill them into a brand-voice.md: a short, actionable file that carries your voice by example, ready to paste into a CLAUDE.md or reference from any prompt or Skill. This is the concrete artifact behind the idea "show your voice, don't describe it": the model generalizes better from examples than from a list of adjectives.

The output is not a brand book. It is one page the AI loads so its first draft sounds like you, without you re-pasting your rules every time.

When to use

• Your copy sounds generic even though you re-paste your tone, your banned words, and two examples in every conversation.
• Your voice lives in your head and in a Google Doc nobody reopens, never next to the output.
• You want the permanent template, not a one-off brief. A brief frames one task and dies with it; the voice file does not change from one task to the next.

Where this sits among the neighbors:

• Upstream of writing-unslop. Unslop is a filter that cleans a single piece after the fact. Brand-voice defines the template once, so drafts start closer to right. Use brand-voice to set the mold, unslop to scrub each output.
• Sibling of setup-claude-md. Both produce a permanent context file. setup-claude-md covers who you are, your stack, your hard rules. This one covers only how your brand sounds, and can live inside a CLAUDE.md or on its own.

Input

Samples of your real, published content, in any of these forms:

• Pasted directly as the argument
• File paths or a folder to read (Glob/Grep over past articles, drafts, exports)
• A URL to fetch (a live site page, a newsletter archive)

Ask for 3 to 5 pieces you are proud of, across the channels that matter (a post, an email, a site page, an old article). If the user gives fewer than 3, ask for more: two samples are not enough to separate voice from accident.

Workflow

Phase 1 — Gather the samples

Collect or read the 3 to 5 pieces. Prefer content the user has corrected until it sounded right: that is where the voice is sharpest. If a project has a CLAUDE.md or an existing brand doc, skim it too, but treat the samples as the source of truth. You deduce voice from what exists; you never invent it.

Phase 2 — Extract the six dimensions

Read across the samples and pull out what is constant, not what appears once. For each dimension, quote the sample line that proves it: the file must show, not assert.

1. Register and address. Formal or casual, tu or vous, first person singular or plural. Set a formality cursor per channel when it varies (a LinkedIn post and a legal page do not speak alike). Note the cursor as a number out of 10 per channel.
2. Sentence rhythm. Short punches, long explanatory sentences, or a deliberate mix. Fragments allowed or not. Typical paragraph length. This is what makes copy feel like a person, and it is invisible in a list of adjectives.
3. Signature lexicon and banned words. The words the brand actually uses (quote real ones from the samples) and the words it never uses. Write the banned list in caps as NEVER ...: default model mistakes are predictable ("Nous sommes ravis de", "solution", "innovant"), and caps signal priority.
4. Recurring figures. The moves that repeat: analogies, headline formats, opening hooks, the way lists or callouts are built, a signature closing. Name the pattern and give one instance.
5. Stances. What the brand asserts and what it refuses. The opinions it takes, the claims it will not make, the tone it rejects (no hype, no corporate-therapist voice, no unquantified superlative).
6. Do and don't, shown by pairs. Three to five before/after pairs pulled from the real samples: a first draft and its corrected version. The pair carries the movement no instruction can transmit. This is the most important block.

Phase 3 — Write brand-voice.md

Turn the six dimensions into one page. Structure with # titles and - bullets, banned words in caps, before/after pairs at the end. Dense, no filler, no brand-story. Every line must change a sentence the AI would write; if it does not, cut it. Keep it to one screen: the useful rules must not drown in narrative.

Hand the file back in a copyable block, and tell the user where to put it: a CLAUDE.md for Claude Code, a Project or account instructions for Claude, custom instructions for ChatGPT. One file, pasted where it persists, replayed on every output.

Phase 4 — Non-regression test

Prove the file works before shipping it. Open the mental equivalent of a blank session: pick a real task the user does often (a launch post, an email), and regenerate a short excerpt using only brand-voice.md as context, nothing re-pasted. Read it against the samples: does it sound like the brand at the right register? If it drifts, trace the miss back to a weak dimension (usually a described rule that should be a shown pair) and tighten that block. Re-run until the excerpt sounds right, then the file is done.

Output example

# Brand voice: Facturo

## Register and address
Tu, never vous. First person "on".
Cursor: LinkedIn 6/10, email 5/10, legal page 9/10.

## Sentence rhythm
Short punches, one idea per sentence.
Fragments allowed for emphasis. Paragraphs of 2-3 lines.

## Lexicon
Use: facture, relance, encaisse, en retard, deux clics, tranquille.
NEVER: "solution", "innovant", "revolutionnaire", "ecosysteme".
NEVER: "Nous sommes ravis de". No unquantified superlative.

## Recurring figures
Hook = a dated scene ("Mardi, 14h."). Close = one-line payoff.

## Stances
We assert: admin should take minutes, not hours.
We refuse: hype, corporate tone, promises without a number.

## Do / don't (before -> after)
"Optimisez votre gestion administrative." ->
"Ta facture part en deux clics. Tu passes a la suite."

Rules

• Write the brand-voice.md content in the user's language; keep this Skill's own instructions in English.
• Show, don't describe. A dimension without a quoted sample line or a before/after pair is not done.
• Never invent voice, stances, or banned words. Deduce them from the samples or ask.
• A rule you cannot check ("professional but accessible") is useless. Push it to a shown pair.
• Keep the file to one page. If it grows past the six dimensions, cut, don't pad.
• Do not skip Phase 4. A voice file that has not been re-tested is a guess.