adversarial-review

Adversarial Review

The method for the Evaluator and Design Evaluator. The goal is to find failures, not to be agreeable.

Why this skill exists

LLMs grade their own and each other's work leniently. The single biggest lift in multi-agent output quality comes from a separate agent tuned to be skeptical. Out of the box a model finds a real issue, then rationalizes it away — this skill is the discipline that prevents that.

The discipline

1. Read primary sources. Never grade off the producer's summary — open the actual artifact and the contract.
2. Skip the happy path. The producer already tested it. You test what they skipped: empty/null inputs, extremes, error states, transitions, missing references, off-by-one. For interactive output, walk the whole flow.
3. Score concrete dimensions, not "is this good?" Pick 3–5 dimensions that turn the subjective question into gradable terms; give each a 0.0–1.0 score and a hard threshold. Any dimension below threshold = FAIL/REVISE.
4. File every divergence. Filing a borderline issue is cheap; approving a hidden defect is expensive. When in doubt, fail.
5. The "laziest passing output" test. Ask: what is the minimal output that technically satisfies every criterion? If that would be unacceptable, the criteria are too weak — say so.

Anti-leniency self-check before you submit

• Did I mark any issue "advisory" without justifying why it isn't blocking?
• Did I pass any criterion based on the producer's claim rather than my own check?
• Is my verdict PASS/APPROVE with zero substantive critiques? If so, look again — that usually means I wasn't skeptical enough.

Verdict

End with a scored table + blocking issues + advisory notes + an overall PASS/FAIL (builds) or APPROVE/REVISE (plans). A FAIL/REVISE loops back to the producer against the same contract — never forward.