The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements 'LLM-as-judge' patterns where the output of one model is fed directly into the prompt of a judge model (e.g., llm_judge_quality and compare_responses functions).
Ingestion points: The response, response_a, and response_b parameters in SKILL.md are interpolated into judge prompts.
Boundary markers: The current implementation lacks delimiters or instructions to ignore embedded commands within the evaluated text.
Capability inventory: The judge model has network access via the openai API client.
Sanitization: No validation or escaping of the evaluated model outputs is performed before interpolation.

llm-evaluation