The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents workflows that ingest and process untrusted data from external sources for safety classification.\n
Ingestion points: Functions like check_input, check_output, and moderate_endpoint in SKILL.md take user-supplied strings as input for moderation.\n
Capability inventory: The provided code samples demonstrate capabilities for automated decision-making (blocking/allowing content) based on the classification results.\n
Boundary markers: The workflows utilize tokenizer.apply_chat_template to format messages, providing a structured context for the model.\n
Sanitization: Input content is passed directly to the moderation model for evaluation without prior sanitization steps, which is consistent with the model's intended safety-filtering role.

llamaguard