The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill inherently processes external data by reading documents authored by other agents or users (e.g., files in docs/<feature>/). This creates a surface for indirect prompt injection where malicious instructions inside the reviewed documents could attempt to influence the reviewer's output. However, the skill includes strong limiting instructions—specifying that the agent is a reviewer, not an author, and must never rewrite the document or implement code from it—which significantly mitigates the risk.\n
Ingestion points: Documents in docs/<feature>/, PRINCIPLES.md, AGENTS.md, and CONTRIBUTING.md.\n
Boundary markers: None explicitly specified for input files.\n
Capability inventory: Execution of git fetch and git log to verify branch status; writing architectural notes to .notes.md files.\n
Sanitization: No explicit sanitization of document content before processing is mentioned.

review-doc