SKILL.md
Style Fingerprint Skill
The #1 complaint about AI drafts is "it doesn't sound like me." This skill fixes it at the root: it studies writing the user has actually shipped, extracts the mechanical, imitable features of their voice, and writes a style card into the Brain — after which every brain-aware skill drafts in their register instead of the model's default.
What This Skill Produces
- A style card (~200-300 words, structured) capturing the user's voice as reproducible rules, not adjectives
- Before/after proof: one paragraph rewritten from model-default into the fingerprinted voice, so the user can verify the card works
- The card saved to
brain/knowledge/style.md(with the user's approval), plus a one-line pointer incontext.mdvoice section
Required Inputs
Ask for (if not already provided):
- 3-5 samples the user wrote and shipped — real emails, updates, PRD sections, posts. More samples of the same genre beat variety. Politely reject samples the user merely approved but didn't write — an edited-by-committee doc fingerprints the committee.
- The target register if samples span several (exec formal vs team casual) — or fingerprint each as a named variant
Extraction Method
Analyse mechanics, not impressions. For each dimension, extract a rule an imitator could follow:
| Dimension | What to extract |
|---|---|
| Sentence rhythm | Median sentence length; short-sentence frequency; do they open paragraphs long or punchy? |
| Register & warmth | Contractions? First person singular or plural? Directness of asks ("please could we" vs "let's") |
| Structure habits | Bullets vs prose ratio; headers or none; do they front-load the conclusion (BLUF) or build to it? |
| Signature moves |
