challenge-audit
Installation
SKILL.md
Why this exists
A review report is itself a set of claims. Reviewers invent gaps, misquote sources, miss the same hole collectively, and declare passes early. Every downstream skill (fix, implement, revision rounds) inherits those defects — paying for phantom work or shipping past real holes. This skill spends one audit pass to price the report before spending many passes acting on it.
Division of labor:
| Skill | Unit of work | Question it answers |
|---|---|---|
| fix | one finding | Is this single claim true, and what fixes it? |
| challenge-audit | the whole report | Can I trust this reviewer at all, and what did it miss? |
| challenger (agent) | a design artifact | Where is this spec / ADR itself wrong? |
Run this first when report trust is in doubt; run fix on the findings that survive.
Process
1 — Assemble the three layers
Collect, as paths for cross-reading (if a round's reports live only in a compacted session, consolidate them — findings, classifications, rulings — into one findings document on disk first; the audit needs a cross-examinable artifact, not reconstructed memory):