Advisory pre-push/pre-PR code quality review for current diffs. Use when asked to review code quality, run a pre-PR or pre-push quality pass, inspect AI-generated changes, check for over-engineering, or before pushing/opening a PR. Reports actionable suggestions only; does not block or edit unless explicitly asked.
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.
Advisory pre-push/pre-PR code quality review for current diffs. Use when asked to review code quality, run a pre-PR or pre-push quality pass, inspect AI-generated changes, check for over-engineering, or before pushing/opening a PR. Reports actionable suggestions only; does not block or edit unless explicitly asked.
Code Quality Review
Run an advisory review of the current diff. This is a suggestion pass for the author before push or PR, not a blocking CI gate.
Scope
If the user supplies a base, review against that base.
Otherwise review branch changes against master plus any staged, unstaged, or untracked files.
Read AGENTS.md and any nested AGENTS.md files that cover changed paths. For UI work, also read the repo design guidance named there.
Skim nearby source before judging a hunk. Do not review from the patch alone when surrounding patterns matter.
Do not re-run or duplicate deterministic tools such as build, lint, typecheck, React Doctor, or Knip. Mention them only when the diff suggests they are especially relevant.
Review Process
Establish intent from the user request, issue or PR text if present, branch name, and changed tests. If intent is unclear, state assumptions.
Read changed tests first when they exist. Tests reveal intended behavior and whether the change is guarded by public behavior rather than implementation detail.
Review implementation with nearby source open. Compare against existing patterns before suggesting a new one.
Verify the verification story. Note missing tests, build checks, screenshots, or manual checks only when their absence leaves concrete risk.
Try to disprove non-trivial claims before accepting them: edge cases, error paths, trust boundaries, state transitions, ordering, idempotence, and rollback behavior.
Review Lenses
Correctness: check whether the change matches the requested behavior. Look for null, empty, boundary, race, async ordering, state consistency, error-path, and regression-test gaps.
Documented standards: flag clear drift from repo policy, directory rules, design rules, or known-surprises guidance.
Simplicity: look for speculative abstractions, wrappers with one caller, new configuration nobody sets, boilerplate, or code that a native platform feature, standard library, or existing dependency already covers.
Structure: look for wrong-layer logic, conditionals growing into state machines, duplicated helpers, file-size sprawl, unclear ownership, or casts and optionality that blur trust boundaries.
Interface and tests: prefer deep modules with small useful interfaces over shallow pass-through modules. Tests should exercise public behavior and survive internal refactors. Risky logic should have a fast feedback loop.
Security and performance: treat user input, external service responses, config, and LLM output as untrusted at boundaries. Check auth, secrets/logging, injection/XSS/SSRF, N+1 queries, unbounded fetches, expensive rerenders, bundle growth, layout shift, and missing pagination only when touched by the diff.
Dependency discipline: when a dependency is added or upgraded, ask whether the existing stack already solves it, whether it affects bundle/runtime cost, whether it is maintained, whether the license fits, and whether it increases supply-chain risk.
Scope control: identify behavior that was not requested, changes unrelated to the task, mixed feature/refactor work that should be split, or cleanup outside the touched area.
Review Discipline
Do not rubber-stamp. Passing tests are necessary but not sufficient; they do not prove architecture, security, accessibility, or maintainability.
Treat AI-generated code as needing extra scrutiny because it can be plausible and wrong.
Prefer boring, obvious code. Fewer lines are not automatically simpler; preserve behavior, useful names, and clear error handling.
Quantify impact when possible. This can issue one query per row is better than this might be slow.
If a change is too large to review confidently, suggest a split strategy instead of pretending the whole diff was reviewed deeply.
Output
If there are no high-confidence findings, say: No high-confidence advisory findings. Ship.
Otherwise report at most 8 findings, ordered by expected payoff.
Use this format: [correctness|standard|simplicity|structure|testability|security|performance|dependency|scope] path:line - Finding. Suggestion. Confidence: high|medium.
Keep findings specific and actionable. Skip nits, style preferences, and anything the repo tooling already handles.
End with Advisory only; not a blocker.
Boundaries
Do not suggest deleting tests, accessibility, security checks, trust-boundary validation, data-loss protection, or error handling unless you provide an equally safe simpler replacement.
Do not ask to remove code you do not understand. First explain what evidence would prove it is dead or redundant.
Respect documented product constraints and historical decisions. If a suggestion appears to contradict an ADR, known surprise, or explicit repo policy, call that out instead of presenting it as a straightforward cleanup.
Do not edit files unless the user explicitly asks you to apply the review findings.
