OpenAI 发文纠正会拖累 Astra 的旧提示

Dmitro Cashback @DmitroCP原推发布于 2026.09.14

在 X 上打开原推

OpenAI quietly told everyone their Astra prompts are fighting the model, and published the fix it's a developer blog post from friday. 10 changes, every one of them a thing you're probably doing right now because it worked on GPT-5: 1. cut every skill description to one line that says when to use it. long descriptions get truncated once you have many skills, and Astra stops seeing them clearly. their good example is two sentences 2. turn any multi-workflow skill into a router: a minimal root doc that points to sub-docs and scripts. reading a fat skill burns context and drags in guidance that doesn't apply 3. stop writing recipes. "models have gotten much better at understanding nuance and ambiguity, so overly specific guidance can now hinder results" 4. check who else reads your skills. guidance that helps Sol or Luna "may overconstrain GPT-6 Astra" 5. kill the "before every edit, read architecture.md, database.md and deployment.md" line in AGENTS.md. replace with "use architecture.md for service boundaries, database.md for schema changes, deployment.md when preparing a deployment." their words: pre-reads before every edit are "a great way to burn context and slow work down" 6. delete every instruction telling it to run tests. older models needed the nudge. Astra runs them on its own, and the nudge causes over-testing 7. give it explicit permission for workflows you know are safe, in AGENTS.md. example: "the local tests use disposable fixtures and have no production access. run them, fix failures caused by the requested change, and rerun affected tests without asking for approval at each step" 8. revisit the "ask first" language. it was written for a model with worse judgment. on Astra it makes the model stop where you'd want it to keep going 9. define what done means, up front. Astra "may reach a first implementation and come back for your review while there's still work to do." if done includes running it, inspecting the result and fixing what fails, say so 10. if you want it to keep exploring past a first pass, say what to explore and where to stop the through-line: every rule was compensation for something the last model couldn't do. this model can. the rule is now the bug this matches what Boris Cherny said at YC in july about Claude Code: delete 80% of the prompt, add back only what it repeatedly trips on. two labs, same instruction OpenAI Developers, "Rethinking skills and prompts for GPT-6 Astra". clip via @sairahul1

这条是引用着另一条写的：x.com/i/web/status/2097663941362983168