GPT-6 Astra needs leaner prompts and fewer guardrails, OpenAI recommends

Details

With GPT-6 Astra, more capable models need less hand-holding — a consistent theme in OpenAI's model transition guidance. Provencher's post focuses on three areas where legacy prompt patterns cause problems.\n\nSkill descriptions: Skills are Markdown files with instructions, resources, and scripts. Their names and descriptions go into context so Codex can pick the right skill. Too many skills force Codex to truncate descriptions, stripping information needed for correct selection. Provencher recommends scope descriptions that are short and precise — a Postgres migration skill should only fire for creating or modifying migrations, not general DB work.\n\nAGENTS.md rules: Requiring the model to read multiple documents (architecture.md, database.md, deployment.md) before every change is overkill for routine edits. Astra can determine what it needs. Instead, Provencher recommends pointing selectively: architecture docs for service boundary changes, DB docs for schema changes, deployment notes for shipping.\n\nApproval gates: Requiring explicit approval for safe, low-stakes operations (e.g. running local tests on throwaway data) generates friction with no safety benefit at Astra's capability level. Explicit permissions for trusted operations reduce repeated confirmation requests.\n\nThe underlying principle: detailed step-by-step sequences that compensate for weaker models now slow down Astra, which handles nuance autonomously. Teams running shared skills need to coordinate — what works for one agent may be too restrictive for another running Astra.