Claude code review skills: from diffs to debugging

Claude code review skills give an agent a procedure for judging a change: what to compare it against, which problems to look for and when to stop. Related skills in this category cover test-driven development, debugging and architecture reviews. We reviewed 32: 23 get an install verdict and 9 come with caveats.

Review is also where a skill’s design shows. A weak one tells the agent to “review the code carefully”. A good one fixes the baseline (the diff since a given commit), names the axes (standards and the spec) and says how to handle disagreement between them.

What to look for in Claude code review skills

Start with the baseline. Does the skill review a diff, a branch or the whole codebase? Does it read your repository’s own documented standards, or does it apply generic rules? Then check how it uses subagents. Several skills here split the work among parallel sub-agents, which needs an agent that can spawn them.

Depth matters, but so does restraint. Some of the best skills in this category are 3 to 9 KB and clean, because they are methods, not machinery. Treat large process packs with more caution: they can push branches, install hooks or send your diff to other vendors’ model CLIs.

Security is where packs differ most. Of the 32 skills, 8 are clean, 16 are marked “review” and 8 have broad access. A skill that only reads files and runs git commands is a different risk from one that pre-approves shell access or pushes commits. The security page explains each label. If a skill uses allowed-tools, remember what the Claude Code docs say: it pre-approves tools for that turn and does not restrict the others.

Recommended code review skills

• code-review reviews the diff since a fixed point along two axes, coding standards and the spec, using two parallel sub-agents and a baseline of common code smells. It uses read-only git commands and can fetch the linked issue with gh or glab. It has about 813,000 recorded installs. It is not a security audit.
• tdd is clean: a 3.6 KB test-driven development guide with two supporting documents, built around red-green slices and behavior over implementation. It has over 1.1 million recorded installs.
• diagnosing-bugs is clean and runs a six phase debugging loop: build a failing signal, minimize it, test ranked hypotheses, fix with a regression test. The agent adds tagged temporary logging to your code, so expect it to slow down easy bugs on purpose.
• Superpowers is a 15 skill methodology that clarifies the goal, plans, works through the plan with subagents and verifies. It registers a hook that injects its introduction skill into every session. Status: review. Expect more questions up front.
• Agent Skills by Addy Osmani has 25 lifecycle skills and 9 slash commands (spec, plan, build, test, review, ship), mostly Markdown. It overlaps with Superpowers, so choose one as your main workflow.

For a framework-specific rulebook, vercel-react-best-practices is clean and holds 70 React and Next.js rules in 8 priority bands, each with a wrong and a right example. Note that the parent Vercel pack is a different story: two of its review skills fetch their rules from GitHub on each run, so we label the repository broad access.

For audits that go deeper into vulnerabilities, see security skills. For plans and process, see planning, TDD and debugging. The category page lists all 32 with depth scores, and how we review explains them.