To make Codex follow the same testing and code-review instructions consistently, put repository-wide defaults in AGENTS.md and package specialized workflows as Skills. Then spell out what Codex should inspect, which checks count as evidence, and what it must report when a check cannot run. Test the instructions on representative tasks and refine them based on the results.
Where should reusable Codex instructions live?
Choose the location by the instruction’s scope. Repository guidance sets defaults for work in a project or directory; a Skill packages a workflow that can be reused across tasks. They can complement each other rather than compete.
As an Amazon Associate I earn from qualifying purchases.
| Consideration | AGENTS.md |
Skill |
|---|---|---|
| Best fit | Repository or directory conventions and task-relevant defaults. | A specialized workflow intended for reuse across tasks. |
| Packaging | Plain-text project instructions. | A directory containing a SKILL.md manifest and, when useful, supporting resources. |
| How it is made available | Codex discovers guidance from configuration and repository directories. The CLI guide describes instructions merged through the directory tree, with more-local guidance taking precedence. | Loading depends on the host and API. OpenAI documents different Skill contexts for local execution, hosted or container use with Responses API shell tools, and Agents API sessions. |
| Maintenance focus | Keep rules relevant to the repository and their directory scope; remove or revise rules that no longer help. | Maintain the workflow and its supporting files as a reusable package. |
Keep repository guidance lean: it affects Codex’s work throughout its scope. In OpenAI’s September 11, 2026 article, “Rethinking skills and prompts for GPT-6 Astra,” OpenAI Developers advises: “Because AGENTS.md applies whenever the model works in your repository, you should frequently revisit each instruction and ask yourself whether it’s still needed.” Avoid blanket requirements to read unrelated documentation before every edit. If a local test workflow is safe and appropriate, repository guidance can authorize it specifically.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesWhat should a code-review instruction ask Codex to do?
Define the scope, the risks to inspect, and the shape of the result. OpenAI’s Codex Prompting Guide prioritizes bugs, risks, behavioral regressions, and missing tests. Ask for findings grounded in the diff or affected behavior, rather than unsupported general concerns.
- Identify the changed area or behavior to review.
- Check for bugs, relevant security or operational risks, regressions, and missing tests.
- For each finding, provide concrete evidence and severity.
- If no findings are identified, say so plainly and list remaining risks or test gaps.
Keep criteria relevant to the change. A small documentation edit should not be forced through unrelated security or integration checks simply because the instruction is applied across the repository.
What should a testing instruction specify?
Give Codex a clear verification surface: the appropriate command or test class, the important scenarios, expected behavior, and what to report if a check cannot run. Asking for tests is not proof that they passed or that the change is correct. Distinguish checks that ran and their outcomes from checks that were unavailable or inconclusive.
- Name the specific tests or validation commands that fit the change.
- Describe the scenarios that matter and the expected behavior.
- Require a report of which checks ran and what happened.
- Explain what to report when a check is blocked, unavailable, or inconclusive.
For a broader task, use a review–repair–validate loop: inspect the result, make focused repairs, run the agreed checks, and iterate until the acceptance evidence is met or a concrete blocker remains. Depending on the task, validation may include tests, policy checks, simulations, or human approval. These are possible validation surfaces, not interchangeable guarantees or a universal ranking. Where safety matters, define the human approval boundary explicitly; a passing automated check alone may not settle the question.
Recommended Free Tools
How can you adapt a reusable instruction?
The following is a practical starting point, not an official OpenAI template. Replace the scope, commands, and scenarios with those that fit your project:
For changes in [scope], review for bugs, relevant risks, behavioral regressions, and missing tests. Run [specific validation commands] for [key scenarios]. Report findings with evidence and severity. If no findings are identified, state that and list residual risks or testing gaps. If a check cannot run, say why and what evidence is still needed.Keep repository-specific rules in the relevant AGENTS.md. Put a repeatable, specialized workflow in a Skill’s SKILL.md, adding supporting files only when they make the workflow easier to apply consistently. The right arrangement depends on the project and the host’s Skill-loading mechanism.
How do you tell whether the instructions work?
Try them on a small, representative set of real or safely constructed tasks. Include a straightforward change, a behavioral edge case, and a case where a test gap is known. Check whether Codex follows the requested scope, runs the named validation, catches known issues, supports findings with evidence, and reports limitations. Revise unclear or ineffective instructions, then run the tasks again. This adapts the review–repair–validate–iterate pattern to instruction evaluation; it does not guarantee a particular improvement in review quality or test coverage.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
