Adversarial Review Skill
Use this skill for an independent final challenge after implementation or a
primary review is complete. Always load
review-verification-protocol
before reporting findings. Load code-review only
when the CodeRabbit CLI is the review tool for this change; skip it when
CodeRabbit already reviews the pull request on the hosting service.
Do not use it to implement changes, debug an active unexplained failure, or replace the first focused review of a diff. One independent review per change is enough: a reviewer subagent's review satisfies it, so don't add a second layer unless that review's evidence is missing.
Workflow
- Compare the request, acceptance criteria, diff, tests, documentation, and prior agent claims.
- Inspect the strongest available evidence instead of accepting summaries at face value. Don't re-run checks the report shows passing; run a targeted check only when a finding needs proof.
- Challenge one plausible hidden failure mode, including invalid states, untested boundaries, scope creep, and regressions.
- Report only findings supported by reproducible evidence. Distinguish blockers from follow-ups and explicitly clear disproved concerns.
- Record skipped checks, assumptions, and residual risk before recommending a merge or handoff.
Output
Return a merge recommendation with verified evidence, blockers, follow-ups, skipped validation, and assumptions that still need proof.