The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted content from external sources (WebFetch, Grep, MCP search) which could contain malicious instructions.
Ingestion points: External data enters the agent ecosystem via tool calls like WebFetch and MCP search in a sub-agent context.
Boundary markers: The skill includes a dedicated 'data-not-directive' rule (rule:data-not-directive) which mandates that all fetched content be treated as inert data, preventing the execution of embedded instructions.
Capability inventory: The system can spawn sub-agents, read files, and write to temporary system paths (Write).
Sanitization: All findings are verified by the lead agent through a 'proof-of-work' process that involves grepping for verbatim quotes in temporary files before relaying content to the caller.
[COMMAND_EXECUTION]: The skill utilizes the Write tool to materialize fetched data into temporary files for its verification workflow.
Evidence: Step 5 of the execution protocol specifies writing non-file sources to the system temp directory using names like prime-fetch-verify-<short-id>.txt.
[DYNAMIC_EXECUTION]: The lead agent dynamically constructs Grep commands based on source paths and verbatim quotes provided by the sub-agent to perform its verification check.

prime-fetch