The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for ingesting untrusted data (PDFs, images, and external search results) into LLM context across several recipes.
Ingestion points: Document parsing in references/recipes/document-processing.md and RAG retrieval in references/recipes/rag-oss.md.
Boundary markers: No explicit prompt delimitation or "ignore instructions" markers are demonstrated in the example snippets.
Capability inventory: The recipes describe systems that utilize file system access, network requests to cloud providers, and LLM tool execution.
Sanitization: There is no discussion of input sanitization for extracted document text before it is processed by the model.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing well-known and reputable industry-standard libraries such as vllm, langchain, qdrant-client, and pdfplumber for various LLM engineering tasks.
[COMMAND_EXECUTION]: The recipes include standard system setup commands (e.g., apt install tesseract-ocr-fra) and model serving commands (e.g., python -m vllm.entrypoints.openai.api_server). These are documented as manual setup steps for the user and do not execute automatically.

llm