GPT-6 Astra changes agent prompting: OpenAI's guide to cleaning up Skills and AGENTS.md

AI寒武纪 · wechat · 2026-09-06

Eric Provencher of OpenAI's Codex team says the babysitting prompts agents accumulated over the past year now backfire with GPT-6 Astra, and it's time for a full cleanup of Skills, AGENTS.md, and task prompts.

Skills

Too many skills get compressed/truncated; descriptions must be short and specify exact trigger conditions (e.g., only for DB migrations).
Use progressive disclosure: root file does minimal routing to supporting docs; don't write step-by-step recipes.
Note skills are read by multiple models — instructions written for older models can limit Astra.

AGENTS.md

Remove "read the whole repo first" rules; keep only relevant, updated docs.
Astra runs tests automatically; old nudging instructions cause duplicate test runs.
Grant permissions directly: declare disposable local environments with no production access so the model can run/fix/retest without asking.

Boundaries and pacing

Astra is highly rule-compliant; overly tight boundaries stall it mid-task. Define completion criteria upfront: run, check, fix loops in one instruction.
His tip: feed the original prompt to Codex and let it re-audit all Skills and AGENTS.md in your projects.

Original post →

GPT-6 Astra changes agent prompting: OpenAI's guide to cleaning up Skills and AGENTS.md

More from coding & agent