critic

critic — adversarial self-review

Process

1. Re-read the output as a skeptic, not as its author. Assume the first draft has a bias toward telling a clean story and look for where that bias distorted the content.
2. Hunt for overclaimed conclusions. Does the output state something more strongly than the underlying papers support?
3. Hunt for cherry-picking. Does the set of papers/results chosen support a tidy narrative while omitting contradicting or complicating work that turned up during search?
4. Hunt for coverage gaps. In a fast-moving subfield, is there recent work (last few months) that should have surfaced in literature_review but didn't get checked?
5. Hunt for ideation blind spots. Do any proposed ideas ignore an existing solution that was already found in the same session's literature review?
6. Report findings before finalizing. Surface anything found to the user rather than silently fixing and hiding the fact a problem existed — the user should know what was caught and corrected.

Anti-Rationalization

• "The draft looks solid" — that's exactly the bias this skill exists to catch; a self-satisfied first read is not a critical read.
• "Cherry-picking wasn't intentional" — intent doesn't matter for the output's accuracy; check the actual paper set regardless of intent.
• "No time for another pass" — this is the last gate before the user sees the output; skipping it means overclaims ship uncorrected.

Evidence

A critic pass is complete when: overclaim check, cherry-pick check, and coverage-gap check have each been explicitly performed and their outcome (found issues or none) is reported.

Red Flags

• Skipping straight to "looks good" without checking each of the four hunt categories
• Finding an issue and fixing it silently without telling the user what was wrong
• Never finding any issues across many consecutive reviews (suggests the check isn't actually adversarial)