The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external documentation, user feedback, and examples to generate or update agent instructions. This creates a surface where malicious content in source material could influence the generated skill's behavior.
Ingestion points: references/synthesis-path.md (external sources) and references/iteration-path.md (traces and user feedback).
Boundary markers: The instructions tell the agent to "Apply trust and safety rules when ingesting external content" and "Remove secrets and unrelated personal data," which serve as procedural boundaries.
Capability inventory: The skill is designed to write/update files and execute local validation commands (npm run chain, npm run catalog) within the project environment.
Sanitization: Procedural guidance for removing sensitive data from examples is included in the instructions.

skill-writer