Codex is best for technical work with files. Start in a narrowly scoped project folder, use Ask for approval, describe the desired outcome and checks, and review the changes before accepting them.
Codex can inspect files, edit code, run commands, diagnose problems, and report its validation. Because those actions can change real files, a safe setup matters as much as a good prompt.
Before your first task
- Choose the correct folder. Use a project or copy that contains only the files needed for the task.
- Confirm recovery. Use version control or make a verified backup before significant changes.
- Start with Ask for approval. This lets Codex work in the current workspace and pause before crossing that boundary.
- Remove secrets. Do not place passwords, tokens, private keys, or production credentials in prompts, files, or logs.
- State the acceptance criteria. Say what must work, what must remain unchanged, and which tests or checks should run.
A strong first prompt
Inspect this project and explain its structure before changing anything. Then diagnose [specific problem]. Implement the smallest safe fix. Preserve [important behavior/files]. Run [relevant test or validation] and report: 1. what changed, 2. which files changed, 3. the validation result, and 4. anything you could not verify. Ask before installing software, accessing files outside this folder, using the network, or performing an irreversible action.
Understand permission modes
| Mode | Use | Caution |
|---|---|---|
| Ask for approval | Recommended starting point for most work. | Review requests to reach beyond the workspace or use additional capabilities. |
| Approve for me / Auto-review | Can reduce interruptions while keeping the same workspace boundary. | Automated review can make mistakes. Inspect the task and resulting changes. |
| Full access | Only for a deliberate, authorized task that genuinely requires broad access. | Can edit files beyond the project and use the network without your approval; risk is substantially higher. |
Approval is not a formality. Check the command or action, the exact target, whether data leaves the device, and whether the change can be reversed.
Review the result
- The change is limited to the requested scope.
- Unrelated user work was not overwritten or reverted.
- The reported tests actually ran and passed.
- Error messages and warnings were not hidden.
- Dependencies, licenses, configuration, and security implications were considered.
- No secret or private information appears in the output, log, or commit.
Model and effort
Routine, clearly defined code changes usually do not require Sol. Start with Terra/Medium for everyday development and Luna/Low for mechanical transformations. Escalate when complexity, risk, ambiguity, or failed validation justifies it.
Related guides
- Choose a model, reasoning effort, and speed without wasting your limits
- Use AI responsibly: protect information and verify the result
Official guidance
See OpenAI’s Permissions, Prompting, and Models guides.
Related to