The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external URLs and user-provided files, which may contain malicious instructions designed to subvert agent behavior.
Ingestion points: In skill.md, Phase 1, Steps 1 and 4 ingest user-provided text, files, and content fetched via WebFetch. Additionally, the 'Error Handling' section specifies using web search to supplement short sources.
Boundary markers: The instructions do not define clear delimiters or 'ignore embedded instructions' warnings to separate source content from agent instructions.
Capability inventory: The agents utilize SendMessage for coordination, TaskCreate/TaskUpdate for workflow management, WebFetch for network access, and filesystem operations to write deliverables to the _workspace/ directory.
Sanitization: There are no documented procedures for sanitizing, escaping, or validating the external content before it is processed by the source-analyst or subsequent writing agents.

content-repurposer