Skip to content

OpenAI / ChatGPT Guide Hub

Codex CLI Best Practices — Design, Change, and Verify in Small, Fast Iterations

Principles (4 Points)

  • Plan First: Start with update_plan to break down into 5–7 steps. Update iteratively.
  • Minimal Diff: apply_patch should be equivalent to 1 commit. Avoid unrelated changes.
  • Proximate Verification: Test the "immediate vicinity" of changes → expand (build/link/simple E2E).
  • Transparency: Privilege escalation/destructive operations/external I/O require 1 action = 1 message for prior agreement.

Template (Example)

1) Plan: Implement ○○ (specify boundaries/exceptions)
2) Read: Related files within 200 lines, verify in stages
3) Patch: Apply only changes (minimal comments)
4) Verify: Test/build/static analysis in proximate → broad order
5) Document: Lightweight update of README/AGENTS.md/comments

Example (Small Fix)

Plan: Change home link from .md → / (2 places)
Read: Relevant section of docs/index.md
Patch: Fix href in a tags
Verify: mkdocs build, local navigation check
Document: Specify reason for change in commit message

Common Pitfalls and Avoidance

  • Bloated patches: Don't mix feature addition, cleanup, and style fixes. Split into separate plans.
  • Hiding failures: Show failure logs and immediately propose alternatives/rollback.
  • "While we're at it" fixes: Turn off-topic fixes into TODOs and defer to separate PR/task.

Efficiency Mode (Dangerous Mode)

For maximum development efficiency, you can completely skip approval prompts with the following command:

# Most important: Start interactive session (equivalent to Claude Code's --dangerously-skip-permissions)
codex --dangerously-bypass-approvals-and-sandbox
# → After session starts, all operations are auto-approved
# → Can execute multiple tasks interactively in succession

# Safer alternative (recommended)
codex --full-auto  # Balanced auto-execution session
codex -a never -s workspace-write  # Auto-execute only within workspace

Warning: --dangerously-bypass-approvals-and-sandbox disables all safety mechanisms. Do not use in production; use only in development environments.

Verification Checklist (This Repository Example)

  • MkDocs: mkdocs build --strict (local) / CI is non-blocking → will be stricter in future
  • Links: pytest --check-links docs/ -q (external links summarized in nightly job)
  • Translation pairs: run the JP/EN article pair validation script (AI-Daily-News excluded)

Costs and Limits

  • No additional API charges when using ChatGPT integration. Model/usage limits follow plan and usage conditions.
  • Benchmarks measured locally (wall-clock time, token count, success rate). Record results in Issues for continuous evaluation.