To create a reusable skill in Codex, use $skill-creator to describe the workflow and any tools it should use, or author the skill yourself in a directory containing a required SKILL.md. Keep the skill’s description specific enough to signal when it applies, and put the detailed instructions and optional supporting files in the skill directory. For GPT-6 Astra, concise, focused instructions and selective loading of extra material help keep guidance relevant.
What a Codex skill contains
A skill packages instructions for a repeatable workflow so Codex can consider them when a request matches. The required file is SKILL.md, stored in the skill’s own directory. Additional resources are optional:
references/can hold supporting material such as policies, schemas, examples, or background.scripts/can handle deterministic computation or file processing.assets/can hold templates or other resources used in the output.
OpenAI’s Build skills documentation describes the format and creation approaches. Add support files only when they make the workflow more useful, and explain in SKILL.md when Codex should consult or run them.
Choose a creation route
| Route | Best fit | What to expect |
|---|---|---|
$skill-creator |
You can explain the workflow and the tools it should use in ordinary language. | Start with a natural-language description of the goal; add custom supporting files if the workflow needs them. |
| Manual authoring | You want direct control of the skill directory and its contents. | Create a directory and add a SKILL.md; supporting references, scripts, and assets are optional. |
| Record & Replay | The workflow is stable and easier to demonstrate than to explain. | For eligible users, the ChatGPT desktop app can record a demonstrated workflow while using Codex on macOS. Availability has platform, feature, account, and geographic conditions. |
For the first two routes, the practical choice is between starting from a described goal and writing the files yourself. Record & Replay is a separate, more limited option; check the current Help Center information for account eligibility, Computer Use requirements, and availability. OpenAI’s documentation says initial availability excludes the European Union, Switzerland, and the United Kingdom.
Recommended Free Tools
#1 Best Overall
Build a focused skill
- Choose one recognizable goal. Define the repeatable task the skill should help with. If workflows have different triggers, inputs, or success criteria, separate them rather than combining them into one broad skill.
- Create the skill. In Codex, invoke
$skill-creatorand describe the goal and any relevant tools, or create the skill directory and requiredSKILL.mdyourself. - Write concise metadata. Start
SKILL.mdwith anameand shortdescription. Make the description clearly state the workflow and when Codex should consider using it. It is the activation cue, so avoid wording broad enough to attract unrelated tasks. - Explain how the workflow should run. In the body, specify expected inputs, important steps, the desired output, assumptions not to make, when to ask a question or stop, and which support files to consult.
- Add only useful supporting resources. Put reference material, assets, or scripts in the appropriate directories, then tell Codex when each one matters.
- Check behavior on varied requests. Try direct and indirect requests that should activate the skill, incomplete inputs, requests that should not activate it, and edge cases where the skill must avoid inventing facts or acting without support.
Adapt existing guidance for GPT-6 Astra
OpenAI’s September 11, 2026 guidance on skills and prompts for GPT-6 Astra recommends revisiting existing skills, AGENTS.md files, and task prompts. Focus on the guidance that improves model decisions rather than accumulating generic instructions.
- Make triggers discriminating. Keep skill descriptions as short as possible while still making the intended workflow clear. Overly broad descriptions can activate on irrelevant tasks; too many long descriptions can make skill selection harder.
- Use progressive disclosure. If a skill covers several substantial workflows, keep its root instructions as a minimal router and link to supporting documents or scripts. Load detailed material when the relevant workflow needs it instead of putting every instruction into the entrypoint.
- Remove contradictions and overbroad rules. For project-wide
AGENTS.mdguidance, direct contributors to the relevant documents for the task rather than requiring a full repository map for every change. Consider that Astra-specific project instructions may also be read by contributors using other models. - Define completion for dependent tasks. When a workflow involves several linked steps, state whether the expected result includes implementation, inspection, and fixing failures. A clear finish line can help Astra know how far to take the task.
Reuse and refine the skill
Once the skill exists, invoke or describe the matching workflow in future Codex requests. If it does not activate reliably, tighten its description around the actual trigger; if it activates too broadly, narrow that trigger. If Codex misses a necessary detail, add the specific instruction or supporting reference where it belongs rather than making every request carry the full workflow. Recheck both requests that should and should not use the skill after changing its instructions.
Quick Recap
Rank #2
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.
