Methodology

We read the SKILL.md frontmatter with a safe YAML parser (raw HTML disabled, alias expansion capped) and never execute any repository content. A skill must declare a valid name and description to be listed; invalid files are excluded from browse and indexing.

Categories come from deterministic keyword, path and tool rules over the skill's name, description and location – not the body, which is untrusted. Low-confidence or ambiguous results are flagged for review. A category may be inferred rather than declared.

Static checks look for risky patterns (shell/network/secret access, destructive or remote-execute commands, obfuscation, prompt-injection language) without running anything. They produce evidence and severity, not a verdict. Static checks cannot prove runtime safety, and false positives are expected – always review the source. See Security.

Each listing keeps its source repository, exact path, commit SHA, content hash and fetch time, so a version always resolves to an exact source commit. Displayed dates are the real source/fetch times, never the time you loaded the page.