Coding agents can generate a lot of plausible code faster than humans can audit it. A reviewer skill gives the next agent a repeatable standard instead of asking it to improvise what "review this" means.
Free · open source · SKILL.mdFindings firstBuilt into Critique Chat
critique-review
Code review skill for AI agents.
A compact reviewer protocol for Codex, Claude Code, Cursor, Windsurf, Cline, and other agents that understand agent skills. It teaches the agent to inspect context, verify claims, separate questions from findings, and give a merge verdict without burying the bug under polite summary text. It is free, open source, and preloaded inside Critique Chat for repository review conversations.
Why we made it
Agents need standards, not vibes.
critique-review started as the review discipline we wanted every Critique agent to carry by default: concise enough to load, strict enough to stop rubber-stamp reviews, and portable enough to hand to any coding assistant outside our product.
Generic review prompts often collapse into style notes, praise, or broad warnings. critique-review forces the useful shape: evidence, file line, failure mode, fix direction, checks, and verdict.
Teams already use Codex, Claude Code, Cursor, Cline, Windsurf, and other agents side by side. A portable Markdown skill lets the review standard travel with the team instead of living in one product.
Why this exists
Less approval theater. More evidence.
Findings first
The output starts with concrete bugs and line references, then moves to checks and residual risk.
Review axes
Correctness, security, tests, architecture, performance, and operations stay visible on every review.
Independent pass
When a second agent is used, it gets raw diff context instead of your private chain of thought.
Comparative position
Smaller than a product. Sharper than a prompt.
This is not trying to replace full PR review platforms. Critique itself runs a much larger review control plane with retrieval, model routing, GitHub checks, and Remedy. This skill is the portable review contract underneath: a free Markdown artifact that improves the behavior of any agent that can read instructions before touching a diff.
General code-review skills
Broad quality frameworks are useful for remembering correctness, readability, architecture, security, and performance. critique-review keeps those axes, then tightens the output into merge-oriented findings and verdicts.
Subagent review prompts
Reviewer-subagent workflows are strong when you need an independent second pass. critique-review bakes in the isolation rule: pass raw diff context and requirements, not the main agent conversation or hidden conclusions.
Product-native reviewers
GitHub Copilot, Cursor Bugbot, CodeRabbit, Greptile, Qodo, and similar tools live where the PR is. critique-review is intentionally lighter: a free portable protocol any agent can read before it reviews.
Output shape
A review format that is hard to misunderstand.
The skill pushes agents away from vague advice and toward reproducible claims: severity, file line, failure mode, fix direction, checks, and verdict.
**Findings**
- [P1] Missing tenant filter on review runs
File: app/api/review-runs/route.ts:84
Impact: A user could read runs from another workspace.
Why: The new query filters by run id only.
Fix: Keep the organizationId predicate used by the existing helper.
**Checks**
- pnpm exec tsc --noEmit
- pnpm test:unit
**Verdict**
Request changesDownload .md
Single-file Markdown copy
Open raw Markdown
Static public file
View source
Skill folder on GitHub
Skills run with agent permissions. Read the Markdown before installing third-party skills, then pin or vendor the version your team trusts.