signature

# Signature — capture a team's style, turn it into rules the AI follows A codebase has a *real* style that's rarely written down — people can't list their own conventions, but recognize a violation instantly. This skill surfaces that tacit style and makes it durable: **a correction you give once becomes a permanent rule, not a repeated fix.** Two phases — capture (private, in the Lab), then distil (name-free, into a field pack). > **The rule that never bends: patterns, not people.** Describe what the code does ("this codebase resolves > X by Y"), never rate a person ("so-and-so writes bad code"). The artifact is the pattern; the person is > incidental — and a file that judges a named person is a career-risk if it leaks. Group by contributor > only as a *source of signal*, phrased neutrally. ## Phase 1 — Capture (private, stays in the Lab) 1. **Consent + Lab first.** Ask before scanning a real, possibly-private codebase (and which paths). All output has names, so it lands in `lab/` (gitignored) — if there's no Lab, set one up (`lab` skill) first. 2. **Two sources — the second is the gold:** - **Repeated corrections (highest value).** What people tell the AI over and over — "do X not Y", "always…", "stop doing…", recurring review comments, AI diffs reverted-and-redone. Explicit, human-validated, de-noised. Weight by **frequency** (seen 20× = a rule; once = noise). Read the corrections log if one exists; help start one if not. - **Codebase patterns (l...