The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources during the review process. * Ingestion points: Untrusted data enters the agent context from pull request diffs, linked issues, briefs, and surrounding code files as described in the Panel and Precision modes. * Boundary markers: The skill instructions advocate for blind and neutral prompting to minimize bias, but the document lacks explicit instructions for using delimiters or ignore-commands to wrap the external content. * Capability inventory: The skill has the ability to read project files, access pull request metadata, and post review comments to the repository. * Sanitization: The skill does not define specific sanitization or filtering for the ingested content, though its multi-model agreement threshold and the proof-table verification step act as functional mitigations against malicious instructions.

cross-model-review