AI Skill Catalogs Can Consume Tens of Thousands of Context Tokens
A Juejin AI author scanned 570 SKILL.md files and identified 402 unique skills whose names and descriptions totaled an estimated 45,634 tokens. The author says skill bodies load on demand, while the catalog is presented in each conversation turn, and reduced the measured catalog to 30 frequently used skills at an estimated 7,683 tokens. The figures come from a local character-based estimate rather than a platform tokenizer, and the reported Hermes Agent comparisons were not independently reproduced.
The reproducible scan and accompanying script provide a practical way to measure skill-catalog overhead and organize skills with routers, concise descriptions and selective activation.
一、先说我量到了什么
Hermes Agent 的 issue 区有人贴过一组数:243 个技能,每次调用多 10 到 15K tokens;130 个,大概 4K。
我不太信,就把自己机器扫了一遍。
结果:磁盘上躺着 570 个 SKILL.md,去重之后是 402 个技能。只算每个技能的 name 和 description,92,006 个字符,折合约 45,634 tokens。
注意,这不是技能的正文。正文是"按需加载"的,你不用它就不进上下文。我量的这 4.5 万,是目录——那本告诉你"我这里有什么"的册子。
而册子是每一轮对话都要摊开在桌上的。