The Agent Skills Directory

Data Ingestion Surface: The skill involves the agent reading and analyzing external files to provide feedback. This is a core component of its review functionality. While reading untrusted codebase files is an indirect prompt injection surface, it is a requirement for the skill's stated purpose.
Ingestion points: Documented in SKILL.md, where the agent is instructed to use tools to read files and explore the codebase.
Boundary markers: The instructions utilize structural delimiters like <what_to_review> and <how_to_inspect> to organize the context provided to the subagent.
Capability inventory: The subagent is permitted to perform file-read and directory exploration operations. It is explicitly instructed: 'Do not make direct code changes.'
Sanitization: The skill relies on structured tags and the agent's reasoning to distinguish between its instructions and the code being reviewed.

rubber-duck