给 Astra 瘦身：Skills 和 AGENTS.md 该删了
作者认为为旧模型堆的 Skills、AGENTS.md 和提示脚手架反而拖累 Astra，给出 10 条清理清单，核心是删掉过度指定的步骤。
your GPT-6 Astra setup is probably too bloated a lot of the Skills / AGENTSmd / prompt scaffolding we built for older models is now getting in Astra’s way which is pretty funny because the highest leverage upgrade might be deleting things This is how I’m thinking of it. We stacked long descriptions, full recipes, and “read the whole repo first” rules. That used to help. With Astra it mostly burns context and makes the model hesitate. Easy cleanup checklist (do this today): 1. Audit your skills list. Delete ones you don’t actually use. 2. Shorten every skill description to WHEN to use it. Not an essay. Not “anything database-related.” 3. Cap how many skills sit in one project. Too many and Codex starts shortening the descriptions to make them fit, so the model can’t pick the right one. 4. Turn fat skills into a thin router. Root file points to deeper docs/scripts only when needed. 5. Cut recipe-style step lists. Astra handles nuance better. Over-specifying can hurt. 6. Trim AGENTSmd to what every task still needs. Drop “read the full repo before any edit.” 7. Stop forcing blanket “always run all tests.” Astra checks work more on its own. Give permission for safe local test loops instead. 8. Soften scary “ALWAYS ask first” rules if you’d actually want it to keep going. 9. Define done up front. Implement, run it, inspect, fix failures. Astra can stop after a first pass if you don’t say otherwise. 10. Ask Astra to audit your Skills + AGENTSmd against this article, then review what it flags and decide what should go. Anyone already trimming for Astra? Curious what you’re deleting first
这条是引用着另一条写的：x.com/pvncher/status/2095991462416490862