adversarial-review_skill

Overview

This skill runs an automated, adversarial code review loop using fresh-eyes reviewer subagents to find bugs, security issues, and maintainability problems before committing. It is designed for completed implementations, refactors, and changes affecting critical paths where a tougher review is warranted. The loop repeats up to three rounds, fixing high-severity findings between rounds until only minor nitpicks remain.

How this skill works

The skill collects a diff to review (unstaged/staged changes, specific files/dirs, or a branch PR diff) and gathers project convention files (CLAUDE.md, lint configs, common guide locations) within a token budget. It spawns a new reviewer subagent each round with the conventions and full diff, then applies fixes for Critical and Suggestion findings, logs Nitpicks, and repeats with a fresh reviewer until stopping criteria are met.

When to use it

• After implementation is complete and you want a rigorous pre-commit review
• For features, refactors, or changes touching critical or security-sensitive paths
• When you want a review tougher than your own, with fresh-eyes critique
• When the change is non-trivial and worth the extra review cost
• As the review step in the pipeline: implement → /simplify → /adversarial-review → /finalize → commit

Best practices

• Run only on completed, meaningful changes — avoid trivial edits or docs-only tweaks
• Provide the full diff context: use --pr for branch diffs or paths for targeted reviews
• Limit convention gathering to CLAUDE.md and lint configs (keep to ~5K tokens)
• Apply only Critical and Suggestion fixes automatically; skip reviewer items that conflict with known conventions and note why
• Always spawn a fresh reviewer subagent each round so feedback stays unbiased

Example use cases

• A refactor touching a shared CLI script that multiple machines rely on
• A security-sensitive shell utility that handles credentials or network access
• A multi-file change that restructures init or boot-time scripts across the repo
• A complex bugfix touching concurrency or race-prone areas in shell orchestration
• Final quality gate before merging a feature branch that affects provisioning