I Scanned 40 Public Agent Skills Before Installation. Here’s What the Scorecard Found
说明与用途
开发者现在拥有SkillPreflight,一个在安装前评估代理技能的静态评分卡,突出安全、文档和令牌使用问题。
问题
Developers need a way to assess the safety and quality of agent skills before installing them, as package ecosystems face trust issues.
结果
All 40 scanned skills passed the static analysis; average score 80.8/100, median 85.5/100. Two skills required high-risk review. Token usage varied widely, with median 840.5 tokens and 90th percentile 2,699 tokens. Reliability and maintainability scored lowest, highlighting gaps in documentation and tests.
工作流
- 1. Identify public skills via GitHub search queries for SKILL.md.
- 2. Pin each skill to a specific commit.
- 3. Run SkillPreflight 0.4.0 with default rules on the frozen sample.
- 4. Review the generated JSON/CSV benchmark data.
- 5. Use the scorecard to decide whether to install or request improvements.
证据
All 40 selected skills scanned successfully.
The grade distribution was: A 1, B 29, C 5, D 3, F 2.
SkillPreflight therefore does not claim that an A grade proves safety.
- 记录日期: 2026-09-02T14:27:41.465Z