The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external content including code, documentation, and plans, which constitutes an attack surface.
Ingestion points: The agent processes external files in SKILL.md through various review phases, such as the Impact, Logic, and Removal lenses, which involve reading existing code and project history.
Boundary markers: The skill includes a 'Handing a review to an outside model' brief that explicitly instructs reviewers to use read-only commands and avoid editing files, which serves as a safety guideline, though no specific data delimiters (e.g., XML tags or unique markers) are mandated for the ingested data.
Capability inventory: The skill utilizes Read, Grep, Glob, and Bash tools. It specifically instructs the agent to run project-specific tools like type checkers (tsc, pyright, go build) and to perform runtime 'probes' on existing code to verify behavior.
Sanitization: There is no mention of sanitizing or escaping the content retrieved from the files under review before it is interpreted by the agent.

review-lenses