Codex Review
Get a second opinion on the current change or a plan from a different model (Codex), and enforce its verdict mechanically. Codex returns a structured JSON verdict — findings scored by priority (0 critical, 1 important, 2 minor, 3 nit) against sev.schema.json — and a gate script turns that into a pass/fail exit code, so a critical or important finding gets fixed or escalated rather than re-argued by the model that wrote the code. Infra failures (rate limit, model reject, empty output) are their own outcome and never count as a clean pass. Use after writing tests, before a deploy or merge, or to
Install
npx skills add iamneilroberts/claude-skills --skill codex-reviewgit clone --depth 1 https://github.com/iamneilroberts/claude-skills.git /tmp/claude-skills
mkdir -p ~/.claude/skills
cp -r /tmp/claude-skills/skills/codex-review ~/.claude/skills/codex-reviewPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# /codex-review — cross-model external review, mechanically gated Codex is the judge; the diff is the defendant. This session does not get to overrule a CRITICAL/IMPORTANT verdict. ## Before you run: should you, and how hard? This tool assumes the review is *warranted now*. If that isn't obvious, run **`/review`** first — the orchestrator decides whether to review at all, at which gate, and sets the round budget, then dispatches here. Quick self-gate if you skip it: - **Plan/spec exists but little/no code** → review the plan (`--plan`), defer the diff review until after build. Don't diff-review an empty tree. - **Trivial / docs-only / tests-only / mechanical rename** → don't review. Reviewing noise erodes the signal of a real finding. - **Just applied a trivial/mechanical fix** → don't re-review; re-review only substantive change. Before reading any findings, **pre-commit the fix bar**: you will fix CRITICAL/IMPORTANT (a core-user-journey break, data loss, security, a privacy/cost leak) and **file** MINOR/NIT as issues rather than fixing them in-loop. Deciding the bar up front is what stops "one more round" creep. ## The five disciplines 1. **Structured verdict** — Codex emits JSON against `sev.schema.json`, not prose. The gate is computed from `priority`, never eyeballed. 2. **Don't re-judge the verdict** — `codex-review.sh` decides pass/fail via exit code. For BLOCK items you **fix or escalate**; you do **not** argue Codex out of a CRITICAL/IMPORTANT. That
Details
- Skill name
codex-review- Repository
- iamneilroberts/claude-skills
- Path
skills/codex-review/SKILL.md- Repo updated
- 21 August 2026
FAQ
What does the Codex Review skill do?
Get a second opinion on the current change or a plan from a different model (Codex), and enforce its verdict mechanically. Codex returns a structured JSON verdict — findings scored by priority (0 critical, 1 important, 2 minor, 3 nit) against sev.schema.json — and a gate script turns that into a pass/fail exit code, so a critical or important finding gets fixed or escalated rather than re-argued by the model that wrote the code. Infra failures (rate limit, model reject, empty output) are their own outcome and never count as a clean pass. Use after writing tests, before a deploy or merge, or to
How do I install the Codex Review skill?
skills CLI (Claude Code, Cursor, Codex and more): npx skills add iamneilroberts/claude-skills --skill codex-review. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Codex Review popular?
It lives in iamneilroberts/claude-skills, which has 37 GitHub stars and 7 forks.
Related skills
Brag
37iamneilroberts/claude-skills
Memorialize something you just built by generating a self-contained "showcase" page for it — a persuasive, good-looking deep dive: what it is, why it's impressive, how it works — with a screenshot if one can be found or captured, and links to related showcase pages. This is the sales/hype-leaning counterpart to /write-post (which drafts a neutral, no-BS technical post); reach for /brag when you want it to sell. Fast by design, like /idea: ask only what's needed, build the page, don't turn it into a project. Triggers on `/brag <subject>`, "brag about this", "make a showcase page for X", "write
Branch
37iamneilroberts/claude-skills
Create an isolated git worktree for a task and log it to a shared work journal, so multiple Claude Code sessions run in parallel without racing on HEAD, uncommitted edits, or the index. Subcommands: `list`, `done <slug>`, `update` (heartbeat), `coord <message>` / `coord clear <slug>` (cross-session constraints). Triggers on `/branch`, "start in a worktree", "isolate this work", "parallel session", "heartbeat the journal".
Claude Code Best Practices
37iamneilroberts/claude-skills
Use when asked about Claude Code workflow, prompt caching cost, session economics, CLAUDE.md hygiene, context-window management, plan mode, permission modes, hooks, skills, subagents, checkpointing/rewind, parallel sessions, worktrees, headless/CI usage, the Agent SDK, slash commands, MCP configuration, or "what's the right way to use Claude Code for X". Also use proactively BEFORE suggesting /model, /compact, fast-mode toggle, or editing CLAUDE.md mid-session, since those bust the system cache.
Comment Refactor
37iamneilroberts/claude-skills
Refactor the comments in AI-generated (or long-lived) code — strip the accumulation of fix-history narration, obsolete explanations, and code-restating noise while preserving load-bearing WHY/TODO/warning comments. Works over a path, glob, or the whole repo; skips dead/abandoned/unused and vendored/generated files; and stamps each cleaned file with a content-hash marker so re-runs skip files that haven't changed. Incorporates the /remove-comments and /deslop (humanizer) heuristics plus a used-code filter and incremental idempotency. Triggers on `/comment-refactor [path|glob]`, "refactor the co
Codex
37iamneilroberts/claude-skills
Delegate a task or question to the OpenAI Codex CLI running as an isolated subagent, then relay its answer. Use when the user invokes /codex <prompt>, says "ask codex", "have codex look at / do X", or wants a second independent model to investigate a question or make a self-contained edit. Read-only by default; /codex --write lets Codex edit the working tree. Distinct from /codex-review, which is the structured JSON review gate — /codex is a general-purpose Codex delegate.
Crisp
37iamneilroberts/claude-skills
Use this skill when the user wants shorter, more direct responses. Activate on: "crisp mode", "go crisp", "/crisp", "be brief", "less tokens", "cut the fluff", "shorter answers", "stop being verbose", "drop the filler", "skip the pleasantries", "no fluff", "just the answer". Do NOT activate for: "explain in detail", "walk me through", "full report", or any question answerable correctly in one sentence without the skill.
Browse the most popular Claude skills, all software development skills, or write your own with the skills generator.