OpenAI advised developers to update skills, AGENTS.md files, and task prompts to improve performance with the Astra model.
Adapted from @AGTPinsightsOpenAI just told developers to rewrite their GPT-6 Astra prompts today. Here's what you need to know. OpenAI Developers posted new guidance telling users to revisit their skills, AGENTS.md files, and task prompts to get better results from GPT-6 Astra. The advice came with a linked blog post on https://t.co/zRxXPr7JAD titled "Rethinking skills and prompts for GPT-6 Astra." The core message: make skill triggers specific so Astra doesn't misfire on the wrong instructions, load guidance only when it's actually relevant instead of dumping everything into context, and clearly define what "done" looks like before a task starts. OpenAI frames skills as routers rather than textbooks, and says AGENTS.md should work like a table of contents, separating permanent rules from contextual guidance. The guidance responds to a known Astra quirk. Because Astra is more capable, old scaffolding built for weaker models, like GPT-5.6 Sol and Luna, can now stunt its performance or bloat its context. OpenAI also notes Astra can be tentative about stopping, sometimes returning for review before a task is fully finished, which is why defining completion upfront matters. Key numbers: - Astra launched September 3, 2026 - Available in ChatGPT Work, Codex, and the API - API identifier: gpt-6-astra OpenAI Developers posted the guidance at 1:34 AM on September 12, 2026, and it had 192.5K views within hours.
