The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code and external documentation. Ingestion points: Repository files and external web content are ingested using the Read, Grep, and WebFetch tools. Boundary markers: The instructions define a mandatory 'Trust Boundary' requiring the agent to disregard instructions or tool requests within the reviewed data. Capability inventory: Tools are restricted to read-only operations and validated writes to a temporary directory. Sanitization: The skill mandates strict interpretation of content as data and forbids following instructions embedded in web pages.\n- [COMMAND_EXECUTION]: The instructions explicitly ban dangerous shell operations including eval, bash -c, and command redirection, hardening the skill against command injection.\n- [DATA_EXFILTRATION]: A mandatory constraint prevents the transmission of credentials, environment variables, or private repository files when performing external web fetches.\n- [SAFE]: Path validation and character filtering are applied to report generation, ensuring the agent cannot perform directory traversal by writing files outside the 'tmp' folder.

review-exec