The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a susceptibility to indirect prompt injection because it reads and processes external artifacts that may contain instructions designed to influence the agent's behavior.
Ingestion points: The skill reads user artifacts using the 'file-read' capability as defined in the open-design.json configuration.
Boundary markers: No specific delimiters or safety instructions are present to differentiate between the artifact data and the skill's internal logic or instructions.
Capability inventory: The skill utilizes 'fs:read' and 'fs:write' capabilities to modify local files, alongside a 'patch-edit' pipeline stage that processes content directly from artifacts.
Sanitization: The skill does not perform validation or sanitization on the content of the artifacts before they are ingested into the refinement loop.

refine-critique-loop