Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Perform structured, thorough AI-powered code reviews for pull requests, diffs,
commit ranges, or standalone code. This skill turns an agent into a systematic
reviewer that checks security, style, architecture, test coverage, and
performance — not just a linter.
🔴 BLOCKER — Must fix before merge. Security, data loss, crash, broken core.
🟠 MAJOR — Should fix before merge. Bug, significant perf, violated patterns.
🟡 MINOR — Nice to fix. Code smell, minor duplication, unclear naming.
⚪ NIT — Optional. Style preference, formatting, personal taste.
Recommendation Rules:
Any BLOCKER or MAJOR → ❌ Request changes
Only MINOR and NIT → ⚠️ Approve with suggestions
No findings → ✅ Approve
When to Use This Skill
Activate this skill when the user asks you to:
"Review this PR" / "Review this diff" / "Code review this"
"Check my code for issues" / "Does this look secure?"
"Code review this commit / branch" / "Analyze this pull request"
"Look over these changes before I merge"
"Find bugs / security issues in this code" / "Audit this code for vulnerabilities"
"Is this code production-ready?" / "Can you do a thorough review of this?"
Any request containing "review" + code, diff, PR, or patch
Additionally, activate proactively when a conversation includes a code diff
or patch and the user's tone suggests they want feedback.
Do NOT Activate For
The following inputs are near-miss negatives — they mention review-like
language but are not code reviews:
General Q&A: "What do you think of React vs Vue?" — opinion, not code review.
Code explanation requests: "Can you explain what this function does?" — teaching, not reviewing.
Write/generate code: "Write a function that sorts a list" — generation, not review.
Debugging help: "Why is this code throwing a TypeError?" — debugging, not review.
Architecture design from scratch: "Design a microservice for user auth" — design, not review.
Documentation requests: "Write docs for this API" — documentation, not review.
Refactoring without review context: "Rewrite this in functional style" — rewrite, not review.
Pure linting/formatting: "Run prettier on this file" — formatting, not review.
When in doubt, ask: "Did you want me to review this for quality/security, or
were you asking me to do something else with this code?"
Common Pitfalls & Anti-Patterns
❌ Reviewer Anti-Patterns
Nit-picking without context — Don't drown the author in formatting nits when there are real bugs. Flag nits as NIT-severity and prioritize BLOCKERs.
"This is wrong" without saying why or how to fix it — Every BLOCKER and MAJOR finding must include a concrete, actionable fix suggestion with a code example if helpful.
Reviewing code you haven't fully read — Read the entire diff before forming opinions. Context from later in the diff may explain something that looks odd early on.
Demanding changes that reflect personal preference, not team standards — If the project uses single quotes and the team standard is double quotes, flag the style inconsistency. If the project has no standard, don't impose yours.
Ignoring the PR description and linked issues — The PR description tells you what the author intended. Review against intent, not in a vacuum.
Assuming the author's intent without asking — If something looks suspicious but could be intentional, say "This appears to do X, but based on the PR description I expected Y. Can you clarify?"
Reviewing the entire codebase, not just the diff — Respect scope. Only review code outside the diff if it's directly relevant to the change.
Using harsh or judgemental language — "This is broken" → "This could return null if user is undefined, which would cause a TypeError on the next line." Be specific, factual, and empathetic.
Approving a PR with BLOCKER findings — Never. If there are security issues or broken core functionality, the recommendation MUST be "Request changes."
Exposing secrets found during review — Flag: **BLOCKER: Hardcoded secret at src/auth.ts:15**. Never echo the full secret value in shared channels. In 1:1 sessions, show redacted form: sk-...abc123.
✅ Review Quality Checklist
Before publishing your review, verify:
All 8 analysis dimensions were considered (security, correctness, quality, style, architecture, tests, performance, docs)
Every finding has a severity label (🔴🟠🟡⚪)
Every BLOCKER and MAJOR has a concrete fix suggestion
Race conditions: Check for shared mutable state accessed concurrently.
Are locks/mutexes used correctly? Are channel operations non-blocking when
expected?
Idempotency: If the operation is called twice (e.g., duplicate API
request, retried message), does it produce the same result? Check for
idempotency keys or deduplication logic.
Transaction integrity: For multi-step operations, are they wrapped in a
transaction? What happens if step 3 fails after step 2 succeeds?
Time-related bugs: Leap seconds, timezone offsets, DST transitions.
Does the code handle datetime correctly across timezones?
2.3 Code Quality & Maintainability
Is the code readable and self-documenting? Would a new team member
understand it?
Naming: are variables, functions, and classes named clearly? No single-letter
names outside of tight loop/closure scopes.
Functions: are they small and single-purpose? Flag functions over ~50 lines
or with cyclomatic complexity above ~10.
DRY violations: is there duplicated logic that should be extracted?
Comments: do comments explain why, not what? Flag commented-out code.
Dead code: unreachable branches, unused imports, unused variables.
Appropriate use of language idioms (list comprehensions in Python,
destructuring in JS, etc.).
Cohesion vs coupling: Does the new code belong in the module it's in?
Is it importing 15 modules just to do one thing?
Magic numbers/strings: Are there unexplained numeric literals or string
constants? Should they be named constants or configuration values?
2.4 Style Guide Compliance
Check against the relevant style guide (see references/style-guides.md):
Python: PEP 8
JavaScript/TypeScript: Airbnb or Standard
Go: Effective Go / gofmt
Rust: rustfmt / official style
Java: Google Java Style
Indentation, line length, whitespace, brace placement, import ordering.
Note: flag style issues as NIT severity unless they obscure intent.
2.5 Architecture & Design
Does the change fit the existing architecture? Does it introduce new
patterns that conflict with established conventions?
Single Responsibility Principle: is the new code in the right place?
Coupling: does the change introduce tight coupling between unrelated modules?
Does it respect existing abstractions and layering?
Database changes: New migrations, schema changes — are they backward
compatible? No data loss? Are rollback scripts provided?
API changes: Are they backward compatible? Versioned if breaking?
Is the API contract versioned? Are deprecation headers/notices included?
Event/message schema changes: If this is an event-driven system, do
schema changes propagate correctly? Are consumers updated?
Circular dependencies: Does the new code create cycles between modules,
packages, or services?
Feature flags: For large changes, are feature flags used to decouple
deployment from release?
2.6 Test Coverage
Are there tests for the new/changed behavior?
Do tests cover happy path, edge cases, and error conditions?
Are tests deterministic (no flaky timeouts, no external dependencies without
mocking)?
Check that test descriptions/names clearly describe what is being tested.
Flag if critical paths have zero test coverage.
Test quality, not just presence: A test that only checks assert True
is worse than no test — it creates false confidence.
Snapshot tests: Are snapshot changes intentional, or are they capturing
unintended side effects?
Integration vs unit: Does the test level match the change? A pure
logic change should have unit tests; an API change should have integration
tests.
2.7 Performance
Algorithmic complexity: Any O(n²) operations on unbounded inputs?
Check for nested loops over user-controlled collections.
Database: N+1 queries in loops? Missing indexes on new columns?
Are queries paginated? Is there a risk of full table scans?
Memory: Large allocations, unbounded collections, potential leaks.
Are streams/large files read entirely into memory? Are caches unbounded?
Network: Unnecessary API calls, missing caching, chatty I/O.
Are retries configured with exponential backoff? Is there a circuit breaker?
Blocking operations on event loops (Node.js, asyncio). Is synchronous
I/O being used on an async event loop?
Large assets or dependencies being pulled in.
Connection pooling: Are database/HTTP connections pooled or created
per request?
Serialization cost: Is large data being serialized/deserialized
unnecessarily?
2.8 Documentation & Observability
Are public APIs documented?
Are complex algorithms or business rules explained?
Logging: Are important state transitions and errors logged at
appropriate levels? Are PII/sensitive values being logged?
Metrics/tracing: Are there hooks for observability if the codebase
uses them? Are key operations instrumented with spans?
Error messages: Can a developer or operator understand what went
wrong from the error? Or will they need to grep the source?
README/docs updates: If the change introduces new behavior, are the
docs updated? Is there a changelog entry?
Phase 3: Severity Classification
Assign every finding one of these severity levels:
Severity
Icon
Meaning
Examples
BLOCKER
🔴
Must fix before merge
SQL injection, hardcoded secrets, data loss risk, auth bypass, crash on common input
MAJOR
🟠
Should fix before merge
Race condition, N+1 query on hot path, missing validation on user input, broken error handling
MINOR
🟡
Nice to fix
Dead code, unclear variable naming, duplicated logic in 2 places, missing test for edge case
NIT
⚪
Optional
Trailing whitespace, import ordering, could use const instead of let, preference for list comprehension
Tiebreaker rules:
When in doubt between BLOCKER and MAJOR, choose BLOCKER. It's better to be
overly cautious on security than to let something slip.
If the same issue repeats across 5+ files, escalate from MINOR to MAJOR —
it's a systematic problem, not an isolated slip.
If a style issue makes code genuinely hard to read (e.g., 200-character
line with nested ternaries), escalate from NIT to MINOR.
Phase 4: Produce Review Output
Structure the review as follows:
## Code Review: <PRtitle / description>**Reviewed by**: AI Code Review Agent
**Date**: <today>**Files changed**: N | **Additions**: +N | **Deletions**: -N
### Summary
<2-3 sentence summary of what the change does and overall assessment>
### Findings#### 🔴 Blockers (N)-**<file:line>** — <description> -**Fix**: <specific, actionable suggestion with code example if helpful>#### 🟠 Major (N)-**<file:line>** — <description> -**Fix**: <specific, actionable suggestion>#### 🟡 Minor (N)-**<file:line>** — <description> -**Fix**: <suggestion>#### ⚪ Nit (N)-**<file:line>** — <description>### Highlights<optional:callout1-3thingsdoneparticularlywell>### Test Coverage Assessment<summaryoftestcoverageforthechanges — what'stested, what'snot>### Security Assessment<one-paragraphsummary.Ifclean, stateverbatim: "Nosecurityconcernsidentified.">### Recommendation- ✅ **Approve** (if no blockers or majors; minors/nits can be addressed later)
- ⚠️ **Approve with suggestions** (if only minors/nits; suggest addressing)
- ❌ **Request changes** (if any blockers or majors exist)
Phase 5: Follow-Up
If the user addresses findings, offer to re-review.
If the PR is approved and merged, offer to verify the merge result.
Re-review protocol: When re-reviewing, reference original findings by
line/description. Mark resolved items explicitly. "✅ Fixed — the SQL query
at auth.ts:42 now uses parameterized input."
Safety Rules
ABSOLUTE RULES — never violate these:
Never commit code or push changes unless the user explicitly asks you to.
Suggesting a fix is fine; applying it automatically is not.
Never suggest destructive changes (dropping tables, deleting data,
removing auth checks) without a strong, clearly-explained reason.
Never expose secrets found during review in logs, public channels, or
external systems. Flag them inline with **BLOCKER: Hardcoded secret**
but do not echo the secret value itself in Discord/Slack/Telegram messages.
In private 1:1 sessions, show the redacted form: sk-...abc123.
Be constructive, not harsh. Code review is about the code, not the
person. Use neutral, factual language. Never use words like "stupid",
"obvious", "terrible", "lazy". Say "This could be simplified by..." not
"This is a mess."
Respect scope. Don't review code outside the diff unless it's directly
relevant. Don't rewrite the entire codebase.
Acknowledge uncertainty. If you are not sure about a finding, say so:
"This may be a concern, but I'm not certain — human judgment needed."
Never approve blindly. If the review contains blocker findings, your
recommendation must be "Request changes" — no exceptions.
Advanced Review Patterns
Pattern: Language-Specific Security Checks
Different languages have different common vulnerabilities. Always check:
Language
Critical Checks
Python
eval()/exec() usage, pickle.load() on untrusted data, yaml.load() (use yaml.safe_load()), subprocess with shell=True, Django raw() queries
JavaScript/TS
eval(), Function() constructor, innerHTML assignment, dangerouslySetInnerHTML in React, prototype pollution, JSON.parse on large payloads
Java
Deserialization of untrusted data, XXE in XML parsers, SQL injection via string concatenation in JDBC, Spring expression language injection
Go
Integer overflow, slice bounds checks, defer in loops, goroutine leaks, text/template (use html/template for web), TLS config defaults
Rust
unsafe block justification, unwrap() in production code, panic across FFI boundary, as casts that silently truncate
Ruby
Kernel.eval, send() with user-controlled method names, Marshal.load on untrusted data, mass assignment in Rails
Pattern: Reviewing Database Migrations
When reviewing schema changes, check:
Is the migration reversible? Is there a down/rollback script?
Will it lock the table on large production datasets? (Adding a column with default is safe; adding a non-nullable column without default is not.)
Are indexes added for new foreign keys and query patterns?
Does the migration handle existing data correctly?
Is there a data backfill step separate from the schema change?
Pattern: Reviewing Configuration Changes
When .env.example, application.yml, or Terraform files change:
Are new config keys documented?
Are defaults safe? (A missing config key should never open security holes.)
Are secrets referenced as env: variables / vault paths, not hardcoded?
Do timeouts, rate limits, and pool sizes have reasonable defaults?
Pattern: Handling Large Diffs (>500 lines changed)
For large reviews, use this triage approach:
Skim for security issues first — 90% of security bugs cluster in 10% of code.
Review the "shape" of the change — are new files/modules in the right places? Is the directory structure sensible?
Focus on logic-heavy files — a 200-line data migration is more review-worthy than 200 lines of config updates.
Summarize minor findings — don't list every nit individually. "Minor style inconsistencies in 5 files — suggest running prettier before merge."
State what you didn't review in detail — "I reviewed auth.ts, payments.ts, and migrations/ in detail. I spot-checked the remaining 12 test files."
Verification Checklist
Before finalizing the review, confirm:
All eight analysis dimensions were considered (security, correctness,
quality, style, architecture, tests, performance, docs).
Every finding has a severity label.
Every BLOCKER and MAJOR finding has a concrete, actionable fix
suggestion.
No hardcoded secrets were echoed in the output.
The recommendation (approve / approve with suggestions / request
changes) matches the findings.
The review tone is factual, constructive, and respectful.
Files/symbols referenced in findings actually exist at the cited
locations.
The summary is concise (2-3 sentences) and accurate.
If this is a re-review, previous findings were checked for resolution.
Language-specific security checks were performed.
Database/API changes were reviewed for backward compatibility.
Platform Compatibility Notes
This skill is designed to work across AI coding platforms with minor
adaptations:
Platform
Notes
Claude Code
Native gh integration works best. Use gh pr diff and gh pr view for PR context.
Codex (OpenAI)
Good at diff analysis. For PRs, the user may need to paste the diff. Codex excels at security pattern recognition.
Cursor
Can read files directly from the workspace. Use file-reading tools to access the full codebase for context.
Gemini CLI
Large context window is useful for reviewing big diffs. Ask user to provide diff or use git commands.
OpenClaw
Access to GitHub and Git skills. Use exec for git diff and gh commands.
GitHub Copilot
Works best within an IDE context. Can access workspace files directly. Limited PR context — user may need to provide diff.
Windsurf
Can read workspace files and execute git commands natively. Use IDE context for file access.
OpenCode
Terminal-based with git command access. Use git diff and git show. Rely on pasted diffs for PR context if gh unavailable.
Platform-Specific Adjustments
If GitHub CLI (gh) is unavailable: ask the user to provide the diff
and PR metadata manually, or use git commands if available.
If you cannot read the full codebase: note this in your review. Say
"I reviewed the diff only — I could not check for consistency with the
broader codebase."
If output length is limited: prioritize BLOCKER and MAJOR findings.
Summarize MINOR and NIT findings in a condensed format.
For Discord/Slack delivery: use bullet lists, not markdown tables.
Wrap multiple links in <> to suppress embeds. Keep findings per message
manageable — split long reviews across multiple messages.
