before-after

Before / after — a worked audit

A synthetic walkthrough of the whole loop:

1. weak-CLAUDE.md — an intentionally weak config that looks reasonable (TypeScript, clean code, tests, secrets…).
2. audit-report.md — what /agent-discipline:audit-setup produces against it: 0/100 🔴 — every line is a preference or slogan, not an enforceable rule. Plus per-criterion evidence and concrete diffs ordered by leverage.
3. After — applying templates/CLAUDE.reference.md + templates/settings.reference.json takes it to ~97/100 🟢.

Both numbers are verified: three independent blind auditors scored the weak config 0/0/0 and the reference 97/97/98 — a 0–1 point spread, the reproducibility the standard promises.

The point: the audit doesn't just grade — it hands you the exact diff to the next band, and shows that the single highest-leverage move is usually enforcement (moving rules from prose into the harness), which no amount of better wording can fake.