两次 Jev 调用把 Claude agent 选错技能率从 16.8% 降到 7.3%

kazurai @k4zxbt原推发布于 2026.09.29

在 X 上打开原推

TypeSafe just halved how often a Claude agent loads the wrong skill Hermes cuts every skill description to 60 characters, so lookalikes get mixed up. two Jev calls now sit in front of that pick: 182 skills -> rank all -> reread the top 3 -> suggest one or nothing the whole recipe, animated on one page: > watch 182 skills narrow down to one suggestion > see the pick flip once the full text gets read > steal the 0.30 floor that decides when to suggest nothing 16.8% -> 7.3% wrong skill loads for Claude Haiku 4.5 on Nous Research's Hermes roster. a cheap fix for any agent carrying a big skill folder

这条是引用着另一条写的：「Jev 与 GPT-5.5、Opus 4.8 做 15 轮内容审核一致性测试」