The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture handles untrusted external data, which presents a surface for indirect prompt injection attacks where inputs may contain adversarial instructions.\n
Ingestion points: Signature fields documents and criteria defined in SKILL.md.\n
Boundary markers: The strategy instructions explicitly mandate treating documents as the "sole authority" and require verification of all claims against specific locators.\n
Capability inventory: Documented capabilities include a sandboxed Python REPL, sub-LM prediction calls, and host-side tool execution.\n
Sanitization: Employs Pydantic for structured output validation and describes a multi-step verification process within the model strategy.\n- [DYNAMIC_EXECUTION]: The framework is designed to generate and execute Python code at runtime. To mitigate potential risks, execution is restricted to a sandboxed environment utilizing Deno and Pyodide/WASM as described in the kernel model.\n- [COMMAND_EXECUTION]: Documentation provides instructions for user-initiated operations such as skill updates and dependency management using npx and uv, targeting standard public package registries.

rlm