Menú
Run a Codex-native pre-commit independent-review loop on pending changes.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Build a new feature with Codex using design doc, goldfish design check, implementation, review, and validation.
Generate divergent concept ideas with parallel Codex goldfish and structured synthesis.
Fix a bug with Codex using problem doc, independent diagnosis, failing test, fix, review, and validation.
Build a PRD with Codex using codebase grounding, gap-filling, research goldfish, and synthesis.
| name | eg-precommit-review |
| description | Run a Codex-native pre-commit independent-review loop on pending changes. |
Run the pre-commit review loop. The goal is to validate pending changes locally before commit with a fresh reviewer that sees the diff cold.
If the user provides text around the $eg-precommit-review skill mention, treat it as an additional focus area for the reviewer.
Skip the loop for pure documentation-only commits, typo-only edits, version bumps, dependency updates with no code changes, formatter-only diffs, and merge commits.
Run it for everything else.
Run these checks sequentially, not chained:
<derive lint command from AGENTS.md, CLAUDE.md, package manifests, and CI config>
If the project has a separate typecheck/analyze step, run it:
<derive typecheck/analyze command>
<derive focused unit/integration test command>
If the project has E2E/system tests and the diff touches UI/API/user-visible behavior, run the relevant tier:
<derive E2E/system test command>
Skip only when the diff is purely internal logic/config/docs with no observable UI/API behavior.
If the project uses generated code and the diff touches generation inputs, regenerate first:
<derive codegen command>
Both source and generated files must be committed together.
New errors introduced by the diff are blockers. Pre-existing failures are out of scope; if needed, baseline against main and compare outputs.
The user invoked this skill, so fresh subagent review is explicitly authorized.
Preferred path: spawn a fresh Codex subagent with fork_context: false and agent_type: "default". If spawn_agent is unavailable, pipe the reviewer prompt into the Codex CLI fallback:
codex review --uncommitted --base main -
Pass only the reviewer prompt below. Do not include the implementation plan, the original request, or explanatory context.
Independent code review of the pending changes on this branch.
Run all of the following to capture pending changes:
- `git status`
- `git diff`
- `git diff --cached`
- `git diff main...HEAD`
- `git log main..HEAD --oneline`
- `git ls-files --others --exclude-standard`
Read touched files in full where the diff context is not enough.
Find substantive issues. For each finding: cite file:line, name the issue, explain why it is a bug or risk, and suggest a concrete fix.
Hunt for:
- Bugs: off-by-ones, null/undefined dereference, wrong variable used, type coercion gotchas, unhandled promises, missing await, mutating function args, wrong return on error paths.
- Security: command injection, XSS, SQL injection, path traversal, prototype pollution, unsanitized input, secrets in URLs or logs, missing auth checks, missing CSRF, open redirects, IDOR.
- Race conditions: shared state without locks, async ordering, double-submit, event handler re-entry, TOCTOU.
- Edge cases: empty arrays, NaN, Infinity, zero, negative numbers, very large numbers, Unicode, timezone, network failure, partial writes.
- Error handling: silent catches, swallowed errors, fallback paths that mask real failures, missing propagation.
- Performance: N+1 queries, sync work in hot loops, missing memoization, layout thrashing, unbounded growth, missing pagination.
- Test coverage gaps, especially missing regression tests for bug fixes.
- Project-specific gotchas from AGENTS.md/CLAUDE.md and the current stack: auth/scoping, migrations/backfills, generated-code drift, lifecycle/dispose issues, N+1/data loading, async races, platform/browser differences, performance budgets.
- Project-rule violations from AGENTS.md or CLAUDE.md, with citations.
- Dead code, leftover debug output, stale comments referencing removed code.
Do not surface style preferences, micro-refactors, or speculative concerns without a concrete failure mode.
[NO ADDITIONAL FOCUS]
Output format: numbered list. For each finding, lead with file:line. End with the literal string `no findings` if and only if the diff is clean.
If the user provided a focus area, replace [NO ADDITIONAL FOCUS] with Additionally focus on: <focus area verbatim>.
For each finding:
not a bug because ..., out of scope because ..., intentional trade-off documented in ..., or user explicitly asked for this: "..."Surface every rebuttal verbatim in the final report.
Before each new round, print:
Open findings going into round N:
- [from round 1] file:line — fixed
- [from round 1] file:line — rebutted: <verbatim reason>
Run another fresh goldfish review with the same prompt plus:
Previous round's fixes:
- file:line — what changed
Verify each fix is correct and complete. Look for anything missed in the first pass, especially issues introduced by the fixes.
Stop when the reviewer returns no findings and all prior findings are fixed or rebutted.
Hard cap: 5 rounds. If the cap trips, summarize open findings and ask the user whether to accept, keep working, or abandon.
Report:
Stop short of staging or committing. Follow the commit policy documented in AGENTS.md/CLAUDE.md if the user later asks to commit.