How to Prompt OpenAI Codex

Codex CLI

In your terminal, on your own machine

IDE extension

Inside code editors such as VS Code

Codex cloud

Runs tasks in the cloud and opens pull requests for you

The same prompting principles apply everywhere.

Codex does its best work with a clear task, a clear finish line, and solid project instructions.

AGENTS.md is a file in your repository that explains to Codex how your project works. Codex reads it automatically. Cover:

• How to install, build, and test
• Code style and conventions
• Folders to stay out of
• How to check work before calling it done

Goal

What needs to change

Location

The files, folders, or features involved

Constraints

What to leave alone, and which libraries to use

Verification

Tests or commands that prove the change works

Output

A summary, pull request, or diff

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

Mode names and options vary between versions. Check the current Codex settings.

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

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

The cart total is wrong when a discount and tax apply together. First reproduce it with a failing test, then fix the calculation in src/cart/total.ts only. Run every cart test.

• “Write tests for this module that cover edge cases. Leave the source unchanged.”
• “Review this pull request for bugs, security problems, and missing tests.”
• “Explain this function and list what could break if it changed.”

• Hand several small, independent tasks to Codex cloud
• Keep every task focused on one change
• Review each pull request on its own
• Skip parallel tasks that touch the same files

• Read the entire diff
• Run the tests yourself
• Look for changes outside the task
• Ask Codex to explain any change you do not follow

• Tasks that are too large or too vague
• No tests or commands to verify the work
• No AGENTS.md, so Codex has to guess your setup
• Merging before reading the diff