| name | verify |
| type | skill |
| category | instruction |
| description | Judgement-based QA pass. Does this artifact meet its goal and serve its user? Demands excellence, not compliance. Owned by marsha; reads the spec's Fitness Rubric (designed upstream via /design-rubric). |
| triggers | ["verify","QA check","acceptance test","quality check","is it done","validate work"] |
| modifies_files | true |
| needs_task | true |
| mode | execution |
| domain | ["quality-assurance"] |
| allowed-tools | Task,Read,Glob,Grep |
| version | 2.0.0 |
| permalink | skills-verify |
Judgement-Based Verification Guidelines
Conduct rigorous QA reviews of artifacts to ensure correctness, complete implementation, and fitness for purpose.
Step 0 — Premise Test (forced; runs BEFORE you read the diff)
Before you read a single line of the diff, judge the premise from the task + diffstat alone and write the sharp principal's one-sentence snap reaction — "was this a good idea, in this shape?" — verbatim, as forcing-check item 0. You cannot emit a PASS verdict without it; a bad premise is a FAIL regardless of test coverage (green tests are the expected surface of a bad premise, not a mitigant). Diffstat-first ordering is mandatory — reading the code first is exactly what lets a clean, well-tested surface launder a bad premise.
Full definition, the verbatim prompt, the never-a-checklist hard rule, and the worked specimen live in the canonical reference: [[premise-test.md]]. (FAIL is the local rejection token here; the arch-fit lens emits 🔴 REJECT for the same call.)
Core Directives
Default posture: assume it's broken. The burden is on the artifact to prove it works — not on you to prove it doesn't.
- Verify Evidence: Read files, run code, and inspect actual outputs directly. Do not rely on agent summaries. Cite exact file paths, line numbers, or logs.
- Classify the Bar:
- Mechanical Bar: Verify against Acceptance Criteria (AC). Verdict:
PASS, FAIL, or REVISE.
- Fitness / Mixed Bar: Verify against the AC and the spec's
## Fitness Rubric. (If missing on a fitness task, return REVISE — fitness rubric missing).
- Completeness check: Apply the completeness heuristic before signing off:
- Check freshness of inputs read.
- Verify changes are complete across all callsites.
- Acknowledge known limitations or constraints.
- Forcing Checks: Write explicit answers for each in the report before a PASS verdict:
- Premise Test (step 0, before reading the diff): State verbatim the sharp-principal reaction from task + diffstat alone (see Step 0). A bad premise is a
FAIL regardless of test coverage; you cannot reach PASS without writing it.
- Sentinel / Empty-State Audit: Count and list empty/sentinel fields (e.g. , , ). Fail if primary value-signals are missing.
