The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements evaluation suites and 'LLM-as-judge' patterns that interpolate external data (test inputs, contexts, and model outputs) into prompts.\n
Ingestion points: The EvaluationSuite.evaluate method and several judging functions (llm_judge_quality, compare_responses, evaluate_against_reference) in SKILL.md ingest untrusted data from test cases.\n
Boundary markers: The prompts utilize structural labels (e.g., 'Question:', 'Context:', 'Response:') to delineate inputs, but do not include explicit 'ignore embedded instructions' directives to the evaluating LLM.\n
Capability inventory: The skill uses model.predict and the anthropic library to send data to external LLMs for generation and judgment.\n
Sanitization: There is no evidence of sanitization or escaping of external content before it is interpolated into the evaluation prompts.\n- [EXTERNAL_DOWNLOADS]: The skill incorporates functionality to fetch pre-trained models for evaluation tasks.\n
Evidence: Code snippets reference models such as microsoft/deberta-xlarge-mnli, microsoft/deberta-large-mnli, and facebook/bart-large-mnli through the Hugging Face transformers library. These resources originate from well-known technology organizations.

llm-evaluation