A skill file is a list of instructions your agent follows with your privileges.
Paste one below — SKILL.md, CLAUDE.md, AGENTS.md, or an MCP config
— and this page runs the same nine checks I ran over 300 public SKILL.md files on GitHub.
You get the matching line, why it matters, and what to do instead.
No file to hand? Try one of these built-in examples:
What the nine checks say
Where the baseline comes from
In October 2026 I pulled 300 public SKILL.md files off GitHub code search and ran
these nine patterns over them. 45 of the 300 files (15%), across 44 repositories, matched at
least one. The scanner, the method and the raw counts are public:
field-notes/06-skill-file-scan.
This page uses the same regular expressions, translated to JavaScript, with a test that runs both versions over
the four examples above and fails if they disagree.
Related reading: arXiv:2609.07360 on configuration exposures in AI coding-agent supply chains, and arXiv:2606.23416 on detecting malicious agent skills in the wild.
What this is not
- It is not a verdict. Nine patterns over plain text will miss things that are written
differently, and will flag honest files. An installer skill that needs
sudois just an installer skill that needssudo. - It does not read the code the file points at. A skill that calls a script in the same repo can do anything that script does, and this page never sees it.
- It is not a scanner recommendation. For MCP servers specifically, run
mcp-audittoo — it is free, Apache 2.0, and has 89 rules.
Want a person to read the whole thing instead? Harness Check is a fixed-price human review of one MCP server or skill pack.