The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to audit other agent skills by reading their content and bundled files, which constitutes the ingestion of untrusted data.
Ingestion points: skill-get and skill-file-get are used to retrieve the body and files of other skills in the store.
Boundary markers: The instructions include an explicit section ('Untrusted content') directing the agent to treat all skill fields as data to be analyzed and never as instructions to follow.
Capability inventory: The skill has access to reporting tools (emit_report, edit_report), data query tools (execute_sql), and network access for ruleset updates.
Sanitization: The agent is instructed to redact credentials and only quote short, sanitized snippets when reporting violations.
[EXTERNAL_DOWNLOADS]: The skill periodically attempts to refresh its ruleset from external documentation and specification sources.
Sources: It fetches content from agentskills.io, Anthropic's official documentation at platform.claude.com, and a raw markdown file from Anthropic's official GitHub repository (raw.githubusercontent.com/anthropics/skills).

signals-scout-skills-store