Choose the review surface before installing a Skill
The word “review” can mean a local diff, a GitHub pull request, a repository invariant, or a repeatable analysis procedure. Those needs belong in different places. Start with the work you need reviewed and choose the smallest mechanism that covers it.
| Need | Start here | Why it fits |
|---|---|---|
| Uncommitted, staged, branch, or commit changes | Codex /review | Reviews a selected local change set without changing the working tree |
| A GitHub pull request | Code Review app / connected GitHub review | Keeps the PR diff, comments, checks, and review context together |
| A durable rule specific to this repository | AGENTS.md | Keeps concise project guidance beside the code it governs |
| Formatting, types, or a reproducible invariant | Tests, linters, and CI | Deterministic checks can block a merge rather than merely suggest a finding |
| A repeated, specialized review procedure | A focused custom Skill | Can package a bounded workflow, output contract, and explicit no-write boundary |
What Codex's built-in review already does
For local work in a Git repository, Codex provides a /review flow with scopes such as reviewing uncommitted changes or comparing against a base branch. It reports prioritized findings without modifying the working tree. This is a product capability, not a Skill that you need to install.
For pull requests, the Codex Code Review app can show the PR description, changed files, comments, and checks. Repository-triggered or automatic review depends on connecting the right source account, repository permissions, and review settings. If you only need a second look at a diff, check whether these existing routes already solve the problem before adding another package.
- Is the target a local Git change or a hosted pull request?
- Can the review inspect the exact diff and the expected behavior?
- Does the current workflow return findings without applying changes?
Put project-specific review rules in AGENTS.md
Use AGENTS.md for durable constraints that reviewers repeatedly need to remember: compatibility promises, data boundaries, or an important cross-service invariant. Codex Code Review can apply relevant repository rules and cite them in a finding, so the rule can travel with the code instead of living only in a reviewer’s memory.
Keep each rule narrow: describe the risk, the condition that triggers review, and a safe path forward. A rule such as “preserve this public API field or add a backward-compatible replacement” is more actionable than “check compatibility.” Scope service-specific rules near the service. Put mechanical style and formatting checks in CI instead of making prose compete with deterministic tooling.
- Start with two or three consequential rules, not an exhaustive policy dump.
- Scope each rule to the directories or change types it governs.
- Include a valid safe alternative so the review can distinguish a violation from an intentional exception.
When a separate Codex code review Skill is worth it
A custom Skill earns a place when it adds a reusable procedure that /review and repository rules do not already express—for example, a read-only, defect-first review format that always inspects the complete target diff, checks relevant tests and call sites, and returns only actionable findings with evidence.
OpenAI’s Codex repository includes a sample review-agent Skill with this kind of explicit contract. It is an example to inspect, not a guarantee that every repository should install it. The sample forbids file edits, commits, pushes, posted comments, and further delegation; its value is the review discipline and output boundary, not an extra claim of review accuracy.
- The Skill owns a distinct review task, not a duplicate trigger for the built-in review.
- Its read, write, command, network, and credential access match the task; read-only means the instructions and bundled scripts must both support that boundary.
- Its findings cite a concrete changed path and evidence rather than speculation or style preferences.
Run a small review-quality check before relying on it
A polished SKILL.md is not evidence that the review is useful. Try the same instructions on three small, controlled diffs: one with a consequential defect the rule should catch, one valid counterexample that should not be flagged, and one unrelated change that should not attract irrelevant findings. Compare the actual findings with the intended result.
- The known defect produces a specific, actionable finding tied to changed code.
- A safe exception and an unrelated diff do not produce noisy or invented findings.
- The review continues through the full requested diff rather than stopping after its first issue.
- The output states test gaps and residual risks; “no findings” is not treated as proof that the change is correct.
A Skill is another reviewer—not a merge gate
Codex review can miss defects or raise findings that need context. Verify each claim against the current diff, expected behavior, and tests. Keep required tests, branch protections, and human approval as the controls that decide whether a change may merge.
Bottom line: do not install a Codex code review Skill just because the search phrase contains “Skill.” First use /review for local diffs, Code Review for the connected PR workflow, AGENTS.md for durable project rules, and CI for deterministic checks. Add a custom Skill only when a controlled trial shows that its unique review procedure adds value without widening permissions or creating noise.