GPT-6 Astra changes how you use Codex: OpenAI’s 6 tips on skills, AGENTS.md and prompts

With GPT-6 Astra you should write fewer instructions for Codex. That is the core of the guide OpenAI published in September 2026, Rethinking skills and prompts for GPT-6 Astra: the skills, AGENTS.md files and prompts you have built up over time need a review, because a rule that helped earlier models can make GPT-6 Astra’s output worse.

I read the whole guide. Below are the six tips I think matter most for anyone working with Codex. OpenAI gives many more in the guide.

Why can old instructions now make Codex worse?

Because GPT-6 Astra handles on its own things that used to need explaining. OpenAI writes that overly detailed guidance, written as a step-by-step itinerary, can now limit the result instead of improving it. The same goes for rigid phrasing built for earlier models, such as "always ask before proceeding": with GPT-6 Astra it can stop the model where no stop is needed.

In practice the philosophy changes: fewer instructions, less scaffolding, less micromanagement, more autonomy for the model.

What are the 6 most important tips?

1. Use skills only for specific workflows. Don’t turn every project rule or piece of knowledge into a skill.
2. The fewer skills, the better. Too many skills eat context and make it harder for Codex to pick the right one.
3. Make skill triggers precise. The description must say exactly when the skill applies.
4. Review your old instructions. A rule that helped earlier models can make GPT-6 Astra worse.
5. Trim AGENTS.md. Instructions about things the model already does become noise.
6. Define when the work is done. More autonomy needs a clear finish line.

Why do a few skills work better than many?

Codex reads the descriptions of every skill to decide which one to use, and each description takes up context. OpenAI’s guide adds a concrete detail: when there are too many skills, Codex starts shortening their descriptions to make them fit. The model sees less of each one and chooses worse.

That is why a skill makes sense only for a precise, repeatable workflow. A project convention or general information belongs somewhere else.

How do you write a good skill trigger?

A Codex skill with a vague trigger gets used at the wrong moment. The example is about databases:

• too broad: "use this skill whenever you work with the database";
• precise: "use this skill when you apply a database migration".

With the first, Codex fires the skill every time it touches the database. With the second, only when it is actually needed. The description should stay short, but it must say exactly when the skill applies.

What should you remove from AGENTS.md?

AGENTS.md applies to every task in the repository, so every line weighs on every task. If the file explains to GPT-6 Astra things the model can now do by itself, those instructions become noise: they waste context, and therefore tokens.

OpenAI’s example is about tests. GPT-6 Astra runs tests and checks its own work by itself, so an AGENTS.md that forces it to do so leads to unnecessary testing. The same applies to rules like "read these three documents before every edit": it is better to point to the right document by context, for example the database docs only when the schema changes.

I have seen the same principle on Anthropic’s side: with Claude 5 the rules of context engineering changed too, and instructions written for older models need pruning.

Why do you need to tell GPT-6 Astra when the work is done?

For me this is the most important tip. GPT-6 Astra has more autonomy, and that is exactly why you need to tell it precisely how far to go.

A request like "implement this feature" leaves the finish line open. A request like "implement this feature, test it and make sure there are no errors" closes it. OpenAI’s guide says the same: if the task includes implementing, checking the result and fixing errors, write that into the request. If you ask it to stop for review after the first implementation, Codex stops early, so use that request only when the review is really needed.

Where should you start?

OpenAI suggests not reviewing everything by hand: you can ask GPT-6 Astra itself to audit your skills and your AGENTS.md against these points. I would go one line at a time, asking whether each instruction is still needed. It is the same reasoning behind the advice to delete your CLAUDE.md every 6 months.

If you use Codex every day, your AGENTS.md and your skills are the place to start: a few lines fewer today can give you a sharper agent from your very next task.