How to Prompt OpenAI Codex

Codex CLI

In your terminal, on your machine

IDE extension

Inside editors such as VS Code

Codex cloud

Runs tasks in the cloud and opens pull requests

The same prompting principles work everywhere.

Codex works best with a clear task, a clear finish line, and good project instructions.

AGENTS.md is a file in your repository that tells Codex how your project works. Codex reads it automatically. Include:

• How to install, build, and run tests
• Code style and conventions
• Folders to avoid
• How to check work before finishing

Goal

What should change

Location

Files, folders, or features involved

Constraints

What not to change, libraries to use

Verification

Tests or commands that prove it works

Output

Summary, pull request, or diff

• Start with modes that ask before running commands or editing outside the project
• Give more autonomy only in projects you trust
• Never give it secrets it does not need

Mode names and options change between versions. Check Codex’s current settings.

• Lower effort for small edits and questions
• Higher effort for complex bugs and refactors

Add CSV export to the reports page. Reuse the table data from src/reports/data.ts, follow the existing button styles, and add a test. Run the test suite and summarize the changes.

The total in the cart is wrong when a discount and tax both apply. Reproduce it with a failing test first, then fix the calculation in src/cart/total.ts only. Run all cart tests.

• “Write tests for this module covering edge cases. Do not change the source.”
• “Review this pull request for bugs, security issues, and missing tests.”
• “Explain this function and list what could break if it changes.”

• Send several small, independent tasks to Codex cloud
• Keep each task focused on one change
• Review each pull request separately
• Avoid parallel tasks that touch the same files

• Read the full diff
• Run the tests yourself
• Check for changes outside the task
• Ask Codex to explain any change you do not understand

• Tasks that are too big or vague
• No tests or commands to verify work
• Missing AGENTS.md, so Codex guesses your setup
• Merging without reading the diff