The Agent Skills Directory

[SAFE]: The skill is a utility framework intended for performance measurement and quality assurance. A review of the instructions and the Python implementation in scripts/evaluator.py revealed no malicious patterns such as prompt injection, data exfiltration, or unauthorized command execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted agent outputs for evaluation purposes, presenting a potential attack surface. However, the implementation is safe as it lacks dangerous capabilities and uses robust processing methods.
Ingestion points: Untrusted data enters the evaluator through the output parameter in the evaluate method of AgentEvaluator within scripts/evaluator.py.
Boundary markers: The skill does not implement specific delimiters but relies on structured rubric definitions for scoring.
Capability inventory: The skill scripts contain no file writing, subprocess calls, network operations, or dynamic code execution capabilities.
Sanitization: External content is processed using safe methods, including string normalization (.lower()) and regular expressions for citation detection.

evaluation