umbel: Skill for Claude Code

Audit: Verify Checks

Verify that existing code satisfies a feature's documented checks. Use after /tend discover on mature codebases, after /tend run completes, or anytime you need an honest read on whether a feature actually works. Owns the audit section of feature polyglots and is the only writer of status = "verified".

What this skill assumes

• Skip auditing polyglots with data.persona or data.opportunity set. Catalog descriptions don't have checks or smoke commands — they aren't verified as bets.
• The feature has checks with verification_recipe (output of /tend brainstorm).
• For checks with validates_job set, the bound persona-job's validation clause is the pass criterion — the test must assert against it, not just the recipe.
• The implementation exists on disk and is testable (steps are at least in-progress; ideally done).
• This skill is the ONLY writer of audit.result and the gate to status: "verified". No other skill sets these fields.

Why this skill matters

Run marks steps done when the agent reports completion. That answers "did the code get written?" - not "does it satisfy the check?" Audit closes that gap. Without an audit run, a feature can sit at progress.implementation === 100 with progress.verification === 0 and nobody notices the lie.

What the audit actually asks

A green badge has to mean three things, not one. For every check, the audit answers:

1. Built? — the code the check claims is done is actually done. No TODO / FIXME / XXX / HACK, no throw 'not implemented' / NotImplementedError, no stub body returning a hardcoded placeholder, no large commented-out blocks, no mock standing in for real behaviour on the production path. A // TODO in code labelled "done" is a contradiction this skill exists to catch — flag it whether or not the project tolerates TODOs while building. That tolerance is the builder's policy; surfacing the gap is the auditor's duty, and it is not optional.
2. Proven? — a real test exists, exercises the real unit (not mocks), runs green this audit, discriminates (fails when you break the unit — see negative control), and asserts the measurable outcome (the persona-job validation clause), not a trivial surface like expect(x).toBeDefined().
3. Built well? — the implementation and test meet the project's quality bar (its CLAUDE.md / referenced rubric) plus the universal engineering floor in the Quality rules.