GPT-6 Astra skill guidance puts instruction selection before longer reading lists
OpenAI's September guide warns that broad skill triggers and long catalogs can complicate selection, with implications for teams maintaining shared repository instructions.
Adding more skills to a GPT-6 Astra setup can make the right instructions harder to find. In a September 11 guide, OpenAI says skill names and descriptions remain in context, while long catalogs can be shortened. Broad or overlapping triggers may load guidance a task does not need.
Its proposed fix starts with selection: name the specific task a skill handles, then link deeper material only when relevant. A database migration skill, for example, should not claim every database question. This separates choosing a workflow from reading all of its reference files.
The same guide urges care with repository instructions shared by Astra, Sol and Luna. A rule written for one model may still be read by another; changes should reflect the repository's real end users and tasks.
For teams maintaining skill libraries, the practical question is whether an agent can select a suitable procedure before following it. These are vendor recommendations, not independent evidence of improved reliability.
프리즘코리아 편집국 > Daniel Reed
