brand-voice

name	brand-voice
description	Builds a usable voice profile from real writing samples — do and don't rules each backed by an example, register tiers for different formats, a banned-words list, and before/after calibration rewrites — then saves it so every writer and agent sounds like the brand. Use when copy sounds different in every channel or new writers need more than "keep it punchy".
argument-hint	[paste 5-10 samples that sound like you + 2-3 that don't]

Brand Voice — a profile someone else could actually write from

Most voice guides are three adjectives on a slide — "confident, human, bold" — which describe every brand and help no writer. A usable voice profile is built from evidence: real samples, patterns extracted, every rule paired with an example, and rewrites that show the voice in motion. The test is simple: could a stranger write a passable post from the profile alone?

Inputs

• 5–10 writing samples that sound most like the brand (best posts, founder emails, call transcripts) and 2–3 that sound wrong: $ARGUMENTS
• From marketing-brief.md if present: the existing Voice section to sharpen, the language bank (the customer's register calibrates yours), and any banned phrases already collected.

Do this

1. Audit the samples. Fewer than five good ones? Build the profile anyway, mark it provisional, and name what would firm it up — usually the founder's sent-mail folder or call transcripts, the two places people write and speak like themselves.
2. Extract the patterns across the good samples:
   • Rhythm — sentence length, punctuation habits, how paragraphs breathe.
   • Register — plain or technical, warm or dry, how direct it is willing to be.
   • Moves — how pieces open, how they close, signature turns of phrase.
   • Absences — the words and moves that never appear. Absences define a voice as much as habits do.
3. Write the do/don't rules — eight to twelve of them, every rule paired with a real line from the samples. A rule without an example is an opinion, and writers cannot calibrate to opinions.
4. Define the register tiers: how the voice flexes by format — social at its loosest, email in the middle, sales pages at full pressure — and name what never flexes, the core that makes it the same brand everywhere.
5. Build the banned-words list from the "sounds wrong" pile, plus category clichés and hype words the audience has learned to scroll past.
6. Write three calibration rewrites: an off-voice sentence, the on-voice version, and one line on what changed. These teach faster than any rule.
7. Save the full profile as brand-voice.md in the project root and update the Voice section of marketing-brief.md with the rules, tiers, banned words, and a link to the full profile.

Output

The voice profile — do/don't rules with examples, register tiers, banned words, three calibration rewrites — saved to brand-voice.md and summarised into marketing-brief.md under Voice.

Rules

• Derive the voice from evidence, not aspiration. Document the voice the best samples already have; changing it is a separate, deliberate decision.
• Every rule carries an example from a real sample. No orphan rules.
• Voice is constant, register flexes. If the tiers disagree on fundamentals, the profile has captured two brands.
• Run the stranger test before saving: if the profile alone could not produce a passable post, tighten it until it could.