adversarial-reviewlisted
Install: claude install-skill leinardi/adversarial-review-loop
# Adversarial Review - adversarial-review-loop
Assume change bypass gate until proven otherwise. Find exact command, repo
state, hook event, model response, failure point, or interleaving that make
unreviewed commit, false approval, lost work, or misleading completion. No
praise, no restyle. Zero findings credible only after real attempts to break
changed behavior.
This skill = review procedure. Runtime code, fixed reviewer prompts, Claude
skill metadata, tests, documented user contracts = project authority. Fixed
reviewer prompt supply output contract → follow exact; this skill no replace
that contract.
## 1. Establish the diff
Never review from memory or user description alone. Read real diff, find intent.
| User intent | Command |
| --- | --- |
| "my work", "before I commit", uncommitted changes | `git status --short`, then `git diff HEAD`; inspect every untracked file too |
| staged changes only | `git diff --staged` |
| branch, "this PR", "ready to merge" | determine the default/base branch, then `git diff <base>...HEAD` |
| specific commit range | `git diff <base>..<head>` |
| GitHub PR number | `gh pr view <n>` for intent and metadata, then `gh pr diff <n>` |
Read `git log --oneline` for range plus any linked plan, issue, PR body. Code
that work but implement different contract = finding.
Read every changed file with enough context to know callers and state
transitions. Gate is Python; inspect module boundaries, exception flow
(`hookio`'s `Decided`/`OutputFailur