The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the creation of RAG (Retrieval-Augmented Generation) pipelines and autonomous agents that ingest untrusted data from various external sources.
Ingestion points: The skill references numerous file converters (PDF, HTML, DOCX) in references/09-converters-preprocessors.md and Agent message inputs in references/03-agents.md that process external content into the agent's context.
Boundary markers: While using system prompts, the examples do not demonstrate robust boundary delimiters or strict instructions to ignore embedded instructions in retrieved data.
Capability inventory: The documented agents possess significant capabilities, including tool-calling, math evaluation, and external service interaction (via MCP tools).
Sanitization: The provided RAG patterns do not include explicit steps for sanitizing or validating the content of retrieved documents before prompt interpolation.
[DYNAMIC_EXECUTION]: Documentation provides code examples that utilize dynamic code execution patterns for tool functionality.
Evidence: In references/03-agents.md and references/00-overview.md, a calculator tool is defined using eval(expression, {"__builtins__": {}}). Although it attempts to sandbox the execution by clearing built-ins, the use of eval remains a risky pattern if implemented by users without more rigorous security controls.
[COMMAND_EXECUTION]: The skill includes an executable Python script for maintenance and integrity checking.
Evidence: The script scripts/check-updates.py is provided to allow users to verify reference integrity and check for upstream framework updates. It performs local filesystem scans and initiates network requests to verify version metadata.

haystack