Brand Voice Extractor

---
name: brand-voice-extractor
description: Use when codifying a brand voice from existing writing. Triggers on "brand voice", "tone guide", "writing voice", "voice rules".
---

# Brand Voice Extractor

"Write in my voice" doesn't work as a prompt. The model guesses, often badly. Voice extraction makes the implicit explicit — so any writer (human or AI) can produce on-voice work without channeling a vibe.

## Required inputs

1. **3-5 samples of your strongest writing** — the pieces you'd point to as "this is us"
2. **1-2 samples of writing you'd reject** — useful for "not this" framing
3. **Audience context** — who's reading (different audiences may have different voice)
4. **Format** — what you'll use the voice doc for (writer brief, AI prompt, both)

## Output

### Voice signature (1-page)
- **Sentence rhythm** — average length, variation, where short sentences hit
- **Vocabulary range** — concrete vs. abstract, common vs. specialized, slang allowed?
- **POV** — first person (singular / plural?), second person, third
- **Tense default** — present (immediate) / past (story-like) / mixed
- **What we do** — the moves that recur (analogies, asides, hard-stop sentences, etc.)
- **What we don't do** — phrases banned, structures avoided

### Common moves (with examples from your samples)
The 5-8 specific writing moves that recur. For each:
- The move
- An example from the samples
- When it's appropriate

### The ban list
Phrases / structures explicitly off-brand:
- Buzzwords ("leverage", "synergy", "best-in-class")
- Sentence patterns ("In today's fast-paced world...")
- Tone moves (humble-bragging, fake personality, urgency theater)

### AI prompt template
A reusable preamble that primes the model to write on-voice:

```
You're writing in {our} voice. Match the following rules:
- {rule 1}
- {rule 2}
- {rule 3}

What we do: {moves}
What we don't do: {ban list}

Here are 2 examples in this voice for reference:
<example>{sample 1}</example>
<example>{sample 2}</example>

Now write {the new piece}.
```

### Calibration test
Generate 3 short pieces using the prompt above. Show them to someone who knows your voice. If they can't tell which 2 are AI vs. 1 human-written, the voice doc works.

## Anti-patterns
- Adjective-based voice docs ("our voice is bold, warm, smart") — useless
- 30-page brand books that no one reads
- Voice rules nobody applies because they conflict with the audience's needs

## What to NOT extract
- Topical expertise (that's a different doc)
- Strategy / positioning (different doc again)
- House style (commas, capitalization — that's the style guide, not voice)