OpenAI is advising developers to review accumulated agent instructions when moving projects to GPT-6 Astra. Eric Provencher says broad skill descriptions, blanket document-reading rules and approval restrictions designed for older models can consume context or make the newer model stop before finishing a task.
Skill names and descriptions enter Codex’s context so it can decide which workflow to load. Too many overlapping descriptions may be truncated or trigger irrelevant material. OpenAI recommends short, task-specific scopes, with a main skill file pointing to supplementary documents only when needed. The same principle applies to repository instructions: a typo fix should not automatically require reading architecture, database and deployment guides.
Teams can also explicitly authorize safe local testing and repairs so the agent does not repeatedly request confirmation. Conversely, permissions should remain narrow around production systems and consequential actions. Provencher notes that Astra can stop earlier than GPT-5.6 Sol, making a clear definition of done important. A request should say whether completion requires implementation, execution, checking results and fixing errors. The guidance is model-specific rather than universal; shared instructions must still work for contributors using older or less capable systems.