On 11 September 2026, Eric Provencher published a Codex post on the OpenAI Developers blog titled Rethinking skills and prompts for GPT-6 Astra. He says coding agents needed less handholding as models got more capable, and that people using Codex have piled up instructions to steer older releases. Those instructions take the form of skills, AGENTS.md, and task prompts. With Astra, he writes, it is more important than ever to revisit them.
Skills, in his account, are prompts stored as Markdown files that can also ship with resources and bundled scripts. They are most useful for a specific workflow or certain apps. People now pack a lot of them into projects. Each skill has a name and a description that are loaded into the model's context so it knows when to use them. Many of those descriptions are too long. When there are too many skills, Codex starts shortening the descriptions to fit, so the model sees less of each one and has a harder time choosing. Descriptions can also contradict each other, or over-emphasize when a skill should fire, so the model loads instructions that do not help the task.
They recently updated the $skill-creator skill's guidance. Descriptions should be as short as possible while still making the trigger clear. His bad example is a Postgres migration skill that says to use it when working with databases, queries, models, or persistence, which can fire on any database work. The good version keeps the same first sentence and narrows the trigger to adding or changing a migration, or reviewing its rollout. For skills with several workflows, the root document should be a minimal router that points to supporting docs and scripts, because reading a skill uses context and can pull in guidance that does not apply. Many skills were written as elaborate itineraries. Models, he says, now handle nuance better, so that specificity can hinder results where it once helped. Repository skills also guide other contributors' agents, which may use different models. Guidance that helps Sol or Luna may overconstrain Astra.
AGENTS.md applies whenever the model works in the repository, so each instruction should be checked for whether it is still needed. Telling the model to read architecture.md, database.md, and deployment.md before every edit is, in his example, too much for a typo. Pointing to those files by task is still useful if the docs stay current. Astra, he writes, can work out what it needs to read. Earlier models needed encouragement to run tests. Astra does that on its own, so the same lines can produce unnecessary testing. The model is thorough, but more tentative about how far to take a task. AGENTS.md can give permission for a workflow known to be safe. He offers local tests that use disposable fixtures and have no production access: run them, fix failures caused by the requested change, and rerun affected tests without asking for approval at each step.
People who added strong ask-first language because a previous model acted without permission should look at that wording again. Astra, which he calls OpenAI's most aligned model, "will not perform tasks unless it knows it is safe." If those lines were written to restrain other models, Astra could take them too seriously and stop where the author would have been happy for it to continue.
Readers used to GPT-5.6 Sol continuing for long stretches may find Astra stopping after a first implementation and coming back for review. He wants completion defined before the work starts. If the task includes getting the implementation running, inspecting the result, and fixing what fails, that should be part of the request. A requirement to stop for review after the first implementation will pull the model toward an earlier stopping point. If more exploration is wanted, say what should be explored and where it should stop. A new model is a chance to clean house. He suggests asking Astra to audit against the article rather than reviewing every instruction by hand.
The advice is written for Codex and for GPT-6 Astra. The post does not measure how other models behave under the same leftover instructions, and it does not show a before-and-after of an audit.