The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from the public skills.sh registry, creating a surface for indirect prompt injection.
Ingestion points: The skill ingests third-party data via the npx skills find command and by reading the SKILL.md files of candidate repositories.
Boundary markers: Detailed guidance in references/trust-and-verification.md explicitly instructs the agent to treat registry descriptions and skill bodies as untrusted data, specifically advising to ignore embedded directives such as 'ignore previous instructions'.
Capability inventory: The skill utilizes npx skills add for code installation and employs Bash, Read, Glob, and Grep for file inspection and tool execution.
Sanitization: The skill enforces a triage rubric (evaluating install counts, repository stars, and source reputation) and requires explicit human confirmation before executing installations.
[COMMAND_EXECUTION]: The skill uses the npx package runner to execute the skills CLI for searching and managing skills. It also includes a local utility script scripts/validate-evals.mjs to verify test suite integrity.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading and installing third-party agent skills from GitHub. It contains safety instructions to prioritize trusted sources like official repositories from Vercel and Anthropic and mandates manual inspection of bundled scripts to detect malicious patterns like piped remote execution.

find-skills