The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of external data (user messages and tool outputs) which are subsequently interpreted by Large Language Models.
Ingestion points: The messages array is ingested via the chat() function in SKILL.md and server-side route handlers in references/server-setup.md.
Boundary markers: No explicit prompt delimiters or boundary instructions are included in the code examples, which is typical for library documentation but presents a standard surface for injection attacks.
Capability inventory: The skill supports full agentic cycles and tool execution on both server and client environments (documented in references/tools.md).
Sanitization: Input validation is handled via Zod schemas for tool inputs, providing a layer of structure, though general message content is passed through to adapters as-is.

tanstack-ai