Best code review skills for Claude Code in 2026

The review tag is the noisiest in the directory. Anything whose description mentions "reviewing" is tagged, so the top of the list by installs is design generators: vercel-react-best-practices, design-taste-frontend, sixteen copies of impeccable. We kept only skills whose primary job is to examine code someone else wrote (or the agent itself wrote earlier) and report on it, then read each file for a concrete reporting format and a stated scope.

What it does: when a task or feature is complete, the agent packages the diff, the original requirements and its own summary and dispatches a reviewer (a subagent or a fresh session) instead of declaring success. About 218K installs, 281K stars, updated 2026-09-03, 2 files, 3 KB, no scripts.

When to use it: before every merge. The catch: it assumes the agent can spawn a reviewer; in Claude Code that is a subagent, in Cursor or Codex you will need to open the review yourself. It also expects the rest of the superpowers workflow to exist.

npx skills add obra/superpowers@requesting-code-review -g -y

What it does: fixes the output format. Each finding is one line with file and line, the problem, and the fix, optionally prefixed with a severity (bug, risk, nit). No preamble, no praise. About 333K installs, 103K stars, updated 2026-09-03, 2 files, 2.6 KB.

When to use it: whenever review output goes to a human who has to act on it. The catch: in our crawl the frontmatter description is a single ">" character, so the agent has almost nothing to match against and may not load it automatically. Call it by name, or use it alongside a skill that decides what to look for.

npx skills add JuliusBrussee/caveman@caveman-review -g -y

What it does: a review across separate axes (correctness, security, performance, readability, test coverage, API design) with a severity scale and a rule that the reviewer must say what would change its verdict. About 36K installs, 92K stars, updated 2026-09-03, one file, 20.6 KB.

When to use it: reviewing a pull request of any size when you want breadth. The catch: 20.6 KB per trigger, and breadth means length; pair it with caveman-review to keep the report readable.

npx skills add addyosmani/agent-skills@code-review-and-quality -g -y

What it does: the other side of the exchange. When feedback arrives, the agent must check whether it is technically right before applying it, ask when it is unclear, and push back with evidence when it is wrong, rather than agreeing and silently degrading the code. About 184K installs, one file, 6.2 KB.

When to use it: any time a human or another agent comments on the agent's code. The catch: it makes the agent slower to comply, which is the point, but you will sometimes have to say "yes, do it anyway".

npx skills add obra/superpowers@receiving-code-review -g -y

What it does: reviews front-end code against a maintained list of interface rules (focus states, form semantics, touch targets, motion, dark mode) and reports violations with file and line. About 604K installs, 30.8K stars, updated 2026-08-28, 1.2 KB.

When to use it: the last step before a UI change ships. The catch: it fetches the rule list from the repository at run time, so it needs network access and can change between runs.

npx skills add vercel-labs/agent-skills@web-design-guidelines -g -y

What it does: reads a diff looking only for things to remove: reimplemented standard library functions, dependencies added for one call, abstractions built for a future that has not arrived. About 20K installs, 123K stars, 2.1 KB, updated 2026-08-07.

When to use it: after an agent has written a large feature, because agents over-build. The catch: it only subtracts, so it is a second reviewer, never the only one; and the stars belong to the ponytail project, not this file.

npx skills add DietrichGebert/ponytail@ponytail-review -y

review-animations critiques motion code against a craft bar: is the duration under 300 ms, is the easing right for an entrance versus an exit, can it be interrupted, does it respect reduced-motion. About 138K installs, 35K stars, updated 2026-08-21, 8.1 KB. Use it when a pull request touches transitions. The catch: it assumes CSS or Motion-style code; canvas and WebGL animation are out of scope.

code-review-excellence is about the comments themselves: lead with the most important finding, separate must-fix from suggestion, explain why, offer the alternative. About 28K installs, 13.2 KB, updated 2026-09-01. Use it when the agent's reviews are technically right but nobody acts on them. The catch: it is written for human reviewers and says little about what to look for.

npx skills add emilkowalski/skills@review-animations -g -y
npx skills add wshobson/agents@code-review-excellence -y

The design generators at the top of the tag (vercel-react-best-practices, design-taste-frontend, impeccable and its fifteen duplicate copies) are excellent but they build, they do not review. code-reviewer from Jeffallan (about 5.5K installs) was last updated 2026-08-07 and overlaps code-review-and-quality. open-code-review from alibaba (about 4.1K) has an empty description in our crawl and depends on its own tool. review-pr from code-review-graph needs a knowledge-graph MCP server.

A reviewer that watched the code being written shares its blind spots. In Claude Code, requesting-code-review dispatches a subagent for this reason. In Cursor or Codex, open a new chat, paste the diff and the requirements, and invoke the review skill by name. The combination that works: code-review-and-quality to decide what to look at, caveman-review to decide how to say it, and ponytail-review as a second pass.