Give Codex One Distinct Job

Lesson 1 · Codex as a Complement · ~10 minutes

You already know how to collaborate with a coding agent. The useful first question is not “How do I prompt Codex?” It is: where does a second agent improve the system?

Your win

Use Codex as an independent reviewer of changes produced by you, Kiro, or Claude. A fresh reviewer is less likely to inherit the implementation thread’s assumptions.

The Surface Map

Need immediate, local iteration?  → CLI or IDE
Need visual diffs + worktrees?    → Codex app
Need hosted GitHub delegation?    → Codex web/cloud
Need repeatable shell automation? → codex exec
Need an independent check?        → /review

The CLI, IDE extension, and app share Codex configuration. The app adds parallel threads and built-in worktree support; the CLI adds a terminal-native interactive flow and non-interactive codex exec. Codex web runs hosted work against connected GitHub repositories.

Keep the expanded Codex surface reference nearby when choosing where a task should run.

Run the Complementary Workflow

  1. Open a repository containing a real uncommitted change.
  2. Start Codex in that repository with codex.
  3. Enter /review.
  4. Select Review uncommitted changes.
  5. Accept only findings that cite concrete behavior, risk, or missing verification.

This creates separation of duties: one agent implements; Codex reviews the diff without editing the worktree. Later, you can reverse the roles.

Judge Findings, Not Confidence

A review is useful only when a finding connects evidence to impact. Codex may sound certain about a weak concern, or cautious about a real defect. Evaluate the chain:

  1. Location: does the finding identify the exact changed code?
  2. Trigger: what input, state, or sequence exposes the problem?
  3. Impact: what observable behavior becomes wrong?
  4. Verification: can a focused test or reproduction prove it?
FindingResponse
Concrete trigger and user-visible impactReproduce it, then fix or add a regression test.
Valid risk but no reachable triggerAsk Codex to trace the call path and provide evidence.
Style preference already covered by toolingReject it and rely on the formatter or linter.
Unrelated redesign suggestionRecord separately; keep it out of the current change.
Useful follow-up

Ask: “Show the smallest input that triggers this issue and the test that would fail before the fix.” This converts review prose into falsifiable evidence.

Check Your Understanding

You want Codex to assess Claude’s local uncommitted change without modifying it. Which capability fits?
Correct. /review starts a dedicated local reviewer and leaves the worktree untouched.
Not quite. The requirement is independent, read-only assessment of a local diff, so use /review.

Do This Now

  1. Choose a current repository with an uncommitted change.
  2. Run the local review workflow above.
  3. Classify each finding as proven, plausible, preference, or out of scope.
  4. For one plausible finding, ask Codex for a minimal reproduction or test.
  5. Record one finding you accepted or rejected and the evidence behind that decision.
Primary Source

Codex CLI features documents local review, interactive sessions, codex exec, MCP, cloud tasks, and approval modes.

Questions? Bring back one review finding you accepted or rejected, and explain why. That evidence will determine the next lesson.
← Previous Next →