Rethinking skills and prompts for GPT-6 Astra

What it claims

OpenAI says instructions written for earlier Codex models (GPT-5.6 Sol, GPT-6 Luna) now actively hurt GPT-6 Astra. Two problems drive this. First, over-specified scaffolding costs context and pushes the model toward guidance that doesn't fit the task. Second, Astra behaves differently: it is more cautious and more likely to stop early. It checks its own work without being told, but it treats strongly worded boundaries literally and may return after a first implementation. The page gives no benchmarks; it is guidance only.

How to use it

Details the summary does not carry:

• Too many skills truncate each other. When a project packs in many skills, Codex shortens their descriptions to fit the context. The model then sees less of each one and picks worse. Keep the skill set small as well as each description short. Scope descriptions to the trigger: the page's example is a skill that should fire on a database migration, not on anything that touches a database.
• Use the updated $skill-creator skill to write new skills. OpenAI says it recently changed this skill's guidance to address these failure modes.
• Replace recipe-style skills. Step-by-step itineraries that helped older models can now make results worse.
• Shared repo skills serve mixed models. Other contributors' agents may run Sol or Luna. Guidance tuned for those can overconstrain Astra, so write for every model that will read the file.
• Grant scoped permission in AGENTS.md. This is the page's verbatim example for letting Astra loop on a safe workflow:

  The local tests 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.

• "Stop for review" instructions cause early stops. A requirement to pause after the first implementation pulls Astra toward stopping sooner. Remove it unless you actually need that checkpoint.
• Write the end state into the task. If the task includes running the result, inspecting it and fixing failures, say so. For open-ended exploration, state what to explore and where to stop.
• Let Astra do the cleanup. OpenAI suggests asking Astra to audit your existing skills and AGENTS.md against this article rather than reviewing them by hand.

Limits

The page is vendor guidance with no measurements. Its claims that Astra "does not need" test reminders, and that Astra "will not perform tasks unless it knows it is safe", are assertions, not results. Before loosening boundary language on anything with real side effects, check those claims against your own runs. The page itself says the same instructions can still be right for other models working in the same repo.