| name | pr-review |
| description | Thorough peer code review of a PR or branch — correctness, bugs, logic errors, edge cases, code quality, simplicity, naming, test coverage. Takes a PR number or base branch argument. Pulls PR metadata via gh if available. |
| allowed-tools | Bash, Read, Grep, Glob, AskUserQuestion |
/pr-review — Peer Code Review
Perform a thorough code review as if you are a developer peer-reviewing a pull request for correctness, bugs, quality, and simplicity.
Arguments: $ARGUMENTS
Steps
1. Parse Arguments & Determine Branches
An argument is required. If $ARGUMENTS is empty or missing, stop immediately and tell the user:
Usage: /pr-review <PR-number> or /pr-review <base-branch>
Examples:
/pr-review 42 — review PR #42/pr-review main — review current branch against main
Do NOT proceed without an argument. Do NOT attempt to auto-detect a PR from the current branch.
The argument must be one of:
| Invocation | Behavior |
|---|
/pr-review #123 or /pr-review 123 | Fetch PR #123 via gh. Extract base branch, head branch, and all metadata from the PR. If the head branch is not checked out locally, fetch it. |
/pr-review main | Use main as the base branch. Use the current branch as the head branch. Do NOT look up a PR — just diff against the base branch. |
2. Gather Context
If argument is a PR number:
Fetch PR metadata via gh:
gh pr view <number> --json title,body,number,url,labels,author,reviews,comments,additions,deletions,changedFiles,baseRefName,headRefName
Extract and display:
- PR title, number, URL
- Author
- Base branch and head branch
- Labels
- PR description/body (use this for additional context on intent)
- Stats (additions, deletions, changed files)
- Any existing review comments (to avoid duplicating feedback already given)
If the head branch is not checked out locally:
- Fetch the head branch:
git fetch origin <headRefName>
- Use
origin/<headRefName> as the head ref for diffing (do NOT check it out — stay on the current branch)
Then gather the diff:
git diff $(git merge-base <head-ref> <base-branch>)..<head-ref>
git diff $(git merge-base <head-ref> <base-branch>)..<head-ref> --stat
git log $(git merge-base <head-ref> <base-branch>)..<head-ref> --oneline
Where <head-ref> is origin/<headRefName> if not checked out, or HEAD if the head branch is the current branch.
If argument is a branch name:
Use the argument as the base branch and the current branch as the head. Do NOT call gh or look up any PR.
Gather the diff:
git diff $(git merge-base HEAD <base-branch>)..HEAD
git diff $(git merge-base HEAD <base-branch>)..HEAD --stat
git log $(git merge-base HEAD <base-branch>)..HEAD --oneline
In both cases:
Read the full content of every changed file to understand the complete context around the changes, not just the diff hunks.
3. Understand Project Conventions
Before reviewing, read CLAUDE.md (if it exists) and scan nearby code to understand:
- Language idioms and style conventions
- Error handling patterns
- Testing patterns and frameworks
- Naming conventions
- Project structure and architecture
This context is essential — do not flag things that follow established project conventions.
4. Review: Correctness & Bugs
This is the most important section. Review every change for:
- BLOCK: Logic errors — conditions that are inverted, off-by-one errors, wrong comparisons, incorrect boolean logic
- BLOCK: Null/nil/undefined dereferences — accessing properties on values that could be absent
- BLOCK: Race conditions — shared state accessed without synchronization, TOCTOU bugs
- BLOCK: Resource leaks — opened files/connections/handles that are never closed, missing cleanup in error paths
- BLOCK: Incorrect API usage — calling functions with wrong argument types, wrong argument order, misunderstanding return values
- BLOCK: Missing error handling — fallible operations where errors are ignored or silently swallowed
- BLOCK: Boundary errors — integer overflow, buffer overruns, out-of-bounds access, truncation
- WARN: Unhandled edge cases — empty inputs, zero values, negative numbers, very large inputs, unicode, concurrent modifications
- WARN: Incorrect assumptions documented in comments that don't match the code
5. Review: Code Quality & Clarity
- WARN: Overly complex code — deeply nested conditionals, long functions doing too many things, convoluted control flow that could be simplified
- WARN: Poor naming — variables, functions, or types with misleading or unclear names
- WARN: Code duplication — repeated logic that should be extracted (only if there are 3+ repetitions or the duplication is likely to cause maintenance bugs)
- WARN: Dead code — unreachable branches, unused variables, commented-out code left behind
- WARN: Inconsistency with surrounding code style — different patterns used in the same file or module for the same kind of operation
- NIT: Minor style issues — formatting inconsistencies, import ordering, unnecessary type annotations (only mention if genuinely distracting)
6. Review: Design & Simplicity
- WARN: Over-engineering — abstractions, indirection, or generalization that isn't justified by current requirements
- WARN: Wrong level of abstraction — too much or too little encapsulation for the problem being solved
- WARN: Leaky abstractions — implementation details exposed through interfaces that should hide them
- BLOCK: Breaking API contracts — changes to public interfaces, serialization formats, or database schemas without migration or versioning
- WARN: Missing validation at system boundaries — user input, external API responses, file contents
7. Review: Test Coverage
- WARN: New behavior added without corresponding tests
- WARN: Changed behavior where existing tests were not updated to cover the new logic
- BLOCK: Tests removed or assertions weakened without clear justification
- WARN: Tests that don't actually test what they claim (e.g. asserting on mocks instead of behavior, tautological assertions)
- WARN: Missing edge case tests for complex logic (empty inputs, error paths, boundary values)
8. Review: Security (surface-level)
Flag only clearly visible security issues. For a thorough security review, recommend /review-security.
- BLOCK: Hardcoded secrets, API keys, tokens, or credentials
- BLOCK: SQL injection, command injection, or path traversal from unsanitized input
- BLOCK: Sensitive data logged or exposed in error messages
- WARN: Missing authentication or authorization checks on new endpoints
- WARN: Any other obvious security concern — note it briefly and recommend
/review-security
9. Review: Performance (only if obvious)
Only flag performance issues that are clearly problematic, not speculative optimization opportunities.
- WARN: N+1 queries — database queries inside loops
- WARN: Unbounded growth — collections that grow without limit from external input
- WARN: Unnecessary work — repeated expensive computations that could be cached, loading entire datasets when only a subset is needed
- BLOCK: Algorithmic complexity issues — O(n^2) or worse on potentially large inputs when a better approach is straightforward
10. Output
Structure the output as:
## PR Review: [APPROVE | REQUEST CHANGES | BLOCK]
> PR #<number>: <title> (if PR metadata was available)
> <url>
> Author: <author> | +<additions> -<deletions> | <changedFiles> files
---
### BLOCK
(none — or findings as table below)
| # | Category | Location | Finding | Suggestion |
|---|----------|----------|---------|------------|
| 1 | Category | file:line | What's wrong and why | How to fix it |
---
### WARN
| # | Category | Location | Finding | Suggestion |
|---|----------|----------|---------|------------|
| 1 | Category | file:line | Concern description | Suggested improvement |
---
### NIT
| # | Finding | Location |
|---|---------|----------|
| 1 | Minor issue | file:line |
---
### What looks good
Briefly call out 2-3 things the author did well (good test coverage, clean abstractions, thoughtful error handling, etc.). Peer review should acknowledge good work, not just find problems.
---
### Summary
X BLOCK, Y WARN, Z NIT
Verdict: APPROVE / REQUEST CHANGES / BLOCK
Verdict logic:
- APPROVE: 0 BLOCK, 0 WARN (NITs are fine)
- REQUEST CHANGES: 0 BLOCK, 1+ WARN
- BLOCK: 1+ BLOCK items
11. Offer Next Steps
After the review, suggest relevant follow-up actions:
- If there are security concerns: "Run
/review-security for a thorough security audit."
- If there are production readiness concerns: "Run
/review-prod for a production readiness check."
- If the review is clean: "This looks good to merge into
<base-branch>."
