adversarial-review

Adversarial Review

Overview

Run a panel of independent, hostile reviewers against a change. Each reviewer assumes the code is broken and tries to prove it, from a single narrow angle. Because each runs as a fresh subagent, none of them inherit the main session's reasoning or the author's rationalizations — that is the whole point. A change that "looks fine" to the person who wrote it (or to the assistant that helped write it) gets attacked from 17 directions by reviewers who were never told why it should work.

This is the full panel, and it is token-heavy on purpose: 17 reviewers plus two gate agents, each reading the diff and the surrounding code. If the user asked for a smaller, quicker, lighter, or cheaper review, stop and run the adversarial-review-quick skill instead; for up to about 500 lines, or to re-check the fixes from an earlier review, run adversarial-review-fast (one reviewer). Run this full panel when the user asks for a thorough or complete review, or when no full review has run on substantial work — over 2000 changed lines, more than one subsystem, or a database schema, sign-in, concurrency, or outside-service change. A plan is sized by what it would build: when the change under review is a plan document, count the code its tickets would touch once built, never the plan's own lines — diagrams, examples, and questions make a plan long without making the change large.

Core principle: the orchestrator gathers the change and the brief once, hands each reviewer the raw change, the brief, and a charter, collects findings, a separate verifier confirms each finding is real, and a separate validator confirms each proposed fix actually works — before anything reaches the user.

Companion files — read them first

question-format.md sits next to this file and defines the one shape every end-of-review question takes: the template, what goes in each part, the word cap, the alternatives, and the one-per-message rule. review-snapshot.md sits beside it and defines how to record the state this review looked at, so a later adversarial-review-fast run can review only what changed since. Read both in full the first time this skill runs in a session, before step 1; read review-snapshot.md again at step 1 and question-format.md again at the STOP step. The summaries in those steps are reminders of text you have already read, never a substitute for it.