The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents a pattern for ingesting untrusted user messages into an AI agent context equipped with functional tools.
Ingestion points: Untrusted data enters the agent context through the message field in the data object within the send_chat_message handler in references/ai-serving.md.
Boundary markers: The implementation utilizes a structured BASE_SYSTEM_INSTRUCTION with step-based directives (e.g., ## STEP 1) to guide agent behavior, though it does not demonstrate the use of hard delimiters for user input.
Capability inventory: The skill defines an AIRunner that grants the agent access to tools such as search_records_by_vector, get_record_details, and a purchase tool, enabling both data retrieval and transactional operations.
Sanitization: The provided reference code focuses on orchestration and does not include explicit logic for sanitizing or filtering external content before it is processed by the agent.

litestar-ai-serving