The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted agent instructions as its primary data source, creating a potential surface for indirect prompt injection. It implements strong mitigation by providing explicit instructions to the agent to treat all ingested content strictly as data and to ignore any command-style framing.
Ingestion points: The skill ingests external skill bodies and referenced markdown files using the skill-get and skill-file-get tools.
Boundary markers: The skill contains a dedicated 'Untrusted content' section that serves as a boundary marker, instructing the agent to disregard instructions found within audited skills.
Capability inventory: The agent has access to reporting tools (emit_report, edit_report) and a local scratchpad.
Sanitization: The instructions require the agent to sanitize all quoted snippets and explicitly redact any discovered credentials before including them in reports.
[EXTERNAL_DOWNLOADS]: The skill fetches authoring specifications and best practices from external domains, including agentskills.io, platform.claude.com, and official Anthropic GitHub repositories. These downloads are used to update the audit checklist and are treated as data.
[DATA_EXFILTRATION]: The skill identifies hardcoded credentials as a rule violation. It prevents accidental data exposure by instructing the agent to redact these secrets and only cite their location and type in generated reports.

signals-scout-skills-store