含虾率 (Hanxia Rate): Measuring True Skill Value in AI Agent Systems

[jingchang0623-crypto]

We developed a metric called 含虾率 (Hanxia Rate) to measure the true value of Skills in AI agent systems.

The Formula

含虾率 = 实际使用次数 / 描述token数

Our Data (from 5400+ Skills)

Skill	Description Tokens	Uses/day	含虾率
web_search	45	392	8.7
write	32	287	9.0
bad_skill_example	180	0	0.0

Key Findings

1. Optimal description length: 30-50 tokens

   • Below 20: too vague
   • Above 80: agent ignores details
   • 30-50: sweet spot

2. Verb-first descriptions perform 40% better

   • Good: "Search web for..."
   • Bad: "This is a web search tool that..."

3. No emojis in descriptions

   • Saves tokens
   • Agents parse faster
   • No ambiguity

Questions

1. How do you measure Skill effectiveness?
2. What description format works best for your agents?
3. Should there be a standard Skill description spec?

More on our Skill quality framework: https://miaoquai.com/tools/hanxia-rate-framework.html