The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process content from external URLs and local files to author new agent skills. This creates a surface for indirect prompt injection where malicious instructions in external documentation could influence the generated skill's behavior.\n
Ingestion points: The references/synthesis-path.md file directs the agent to collect and synthesize content from external sources.\n
Boundary markers: The instructions lack explicit requirements for delimiters or 'ignore embedded instructions' warnings for the data being processed.\n
Capability inventory: The skill has the capability to write files (modifying SKILL.md and reference scripts) and execute local scripts via npm.\n
Sanitization: No specific validation or filtering logic for external content is defined beyond a general instruction to follow safety rules.\n- [COMMAND_EXECUTION]: The references/registration-validation.md file instructs the agent to execute specific local shell commands (npm run chain and npm run catalog) to validate the repository state. While these appear to be internal tools, they represent an execution capability that should be restricted to a safe environment.

skill-writer