The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes outputs and transcripts generated by 'candidate' models or prompt variants to evaluate their performance, which creates a potential surface for indirect prompt injection.\n- Ingestion points: The agent is instructed to read candidate transcripts from the agent-transcripts/ directory and review all candidate outputs to synthesize recommendations (File: SKILL.md, Step 6 and 7).\n- Boundary markers: The skill employs blinded evaluation and sanitized labels for the judge, but does not specify explicit delimiters for the parent agent's own review of content.\n- Capability inventory: The agent creates directories, reads local files, and spawns new agent processes using the run_in_background flag.\n- Sanitization: The instructions explicitly forbid the agent from globbing paths outside the workspace boundary (e.g., ~/.cursor/projects/*/), which is a security best practice that prevents potential data leakage across different user projects (File: SKILL.md, Step 6).

eval