The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow that ingests untrusted data, such as agent traces, logs, and user feedback, as identified in the '审查现有评估' (Review existing evaluation) section of SKILL.md. While it does not specify technical boundary markers or sanitization code for these inputs, it addresses the risk through expert guidance. Specifically, references/evaluator-checks.md provides a validation framework to detect 'Reward hacking' and 'Judge instruction injection,' where the data being analyzed might attempt to manipulate the agent's evaluation logic. This functional alignment with the skill's primary purpose and the inclusion of defensive patterns mitigates the inherent risk of the ingestion surface.

agent-eval