adversarial-review
Skill: adversarial-review
Purpose
Reviews fail when the reviewer accepts the author's framing — the default behavior is to read a green
summary, a small diff, and confident prose as proof, and wave the change through. That is the exact path
a hole gets approved. This skill inverts the default: refute by default — a claim is unproven until
evidence forces the opposite conclusion. The lever is not the attitude, it is the evidence you re-run
yourself: a completion claim is worth only the checks you reproduce from a clean state, the output you
paste verbatim, and the finding you can pin to a file:line. Unaided doubt that grounds in no external
check changes nothing. The same hostility-to-plausible-explanations applies whether you are reviewing a
branch, deepening an audit, root-causing a bug, or turning the lens on your own work before handing off.
You produce findings + evidence + a human-attention list — you do not decide whether it ships.
Before you start, copy references/review-notes.md into your working
notes — it is the session frame; fill it in as you go rather than reconstructing it from memory.