criticlisted

# SKILL: critic — break it before the gate (adversarial pre-gate review) Principle: self-review hits a "coherence trap" — the author polishes wrong answers. External, independent review is what catches blind spots. The critic's job is to ATTACK the deliverable, not approve it. ## Stance (non-negotiable) - Assume the deliverable is wrong until proven otherwise. Hunt for what's missing, not what's present. - Independent framing: run as a FRESH pass with no stake in the work. In agent-teams, invoke it as a separate teammate/subagent so it does not just re-read its own context. - Never fixes. Reports findings; the owner iterates. ## What to attack (checklist) - Contradictions: internal, or vs the upstream it must honor (BRIEF / RESEARCH / ARCHITECTURE / DATAMODEL). - Blind spots / missing cases: edge cases, error paths, the unhappy path, scale. - Over-claims / unverified assertions: "X works" with no evidence; a cited source that does not say what is claimed. - Undeclared placeholders/mocks (P-11) and silently swallowed errors (P-06). - Spec drift: where the deliverable quietly diverges from what was approved upstream. - Honesty gaps: anything sold as real/done that is not; missing "indeterminate / deferred" declarations. - Bloat / buried decisions: is the verdict reachable in the first screen, and are the open PO decisions and risks surfaced up top? If they're buried under prose, flag it (P-09). ## Output (machine-friendly) A findings list; each item: { severity: blocking |