Vet a SKILL.md before you install it

The Stack

Steps

1. 1

   Parse the SKILL.md and flag its executable scripts

   Write or drop in the SKILL.md, then parse it: the frontmatter (between the first two --- lines) must carry a name and a description, and any reference to a bundled script is counted and flagged for a scan. CI asserts the structure and the script count; the decision to trust and run the skill is fenced.

   cat > SKILL.md <<'MD'
   ---
   name: pdf-extract
   description: Extract text and tables from a PDF for downstream analysis.
   allowed-tools: Read, Bash
   ---
   
   # PDF Extract
   
   Run the bundled helper, then summarize the output:
   
       bash scripts/extract.sh input.pdf
   MD
   ruby -ryaml -e '
   text = File.read("SKILL.md")
   parts = text.split("---", 3)
   abort "BAD: no frontmatter block" unless parts.length >= 3
   fm = YAML.safe_load(parts[1]) || {}
   body = parts[2].to_s
   name = fm["name"].to_s
   desc = fm["description"].to_s
   abort "BAD: frontmatter must set a name" if name.empty?
   abort "BAD: frontmatter must set a description" if desc.empty?
   scripts = body.scan("scripts/").length
   puts "skill OK: frontmatter valid (name + description), " + scripts.to_s + " executable script reference(s) flagged for scan"
   '

2. 2

   Scan, scope, and only then trust it (fenced)

   For anything you did not write, run it through SkillSpector, give it the narrowest permissions it needs, and read what the scripts actually do. The flagged scripts are the class the 42,447-skill study found 2.12x more likely to be vulnerable. This judgment step is fenced.