The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process and evaluate external LLM outputs, which represents a potential attack surface for indirect prompt injection. \n
Ingestion points: Untrusted data is ingested through fields like input, output, context, and ground_truth defined in the templates/test-case-template.yaml. \n
Boundary markers: The judge prompt templates in reference/llm-as-judge-guide.md utilize Markdown headers (e.g., ## Original Input) to attempt to isolate untrusted data from instructions. \n
Capability inventory: The documented workflow includes LLM invocation and the potential execution of local evaluation scripts. \n
Sanitization: The guides recommend requesting structured JSON output from judge models to facilitate safer programmatic parsing of results. \n- [EXTERNAL_DOWNLOADS]: The implementation guides in reference/llm-as-judge-guide.md provide code examples that utilize standard AI development packages, such as the Anthropic SDK for both Python and Node.js. \n- [COMMAND_EXECUTION]: The skill describes a workflow that involves executing evaluation scripts via the command line, as shown in the SKILL.md examples using the Bash tool to run python evaluate.py.

evaluation