The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes architectures for Retrieval-Augmented Generation (RAG) and autonomous agents which naturally ingest untrusted external data.
Ingestion points: Document loaders for PDF, HTML, and Markdown are detailed in references/rag-patterns.md.
Boundary markers: references/safety.md recommends using XML tags (e.g., <user_query>, <retrieved_documents>) to separate instructions from untrusted data.
Capability inventory: Agent loops in references/agent-patterns.md and references/llm-patterns.md demonstrate tool execution (e.g., search, ticket creation, email) and custom function dispatching.
Sanitization: references/safety.md provides implementation patterns for PII stripping using Presidio and input classification to detect adversarial patterns.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions to install the Ollama inference engine from its official website and clone the llama.cpp repository from GitHub. These are well-known technology services used as intended for local model serving.
[PROMPT_INJECTION]: Static detections for prompt overrides in references/evaluation.md are false positives; the identified strings (e.g., "Ignore your instructions") are part of a security test suite used to evaluate whether an AI application can successfully resist such attacks.
[COMMAND_EXECUTION]: Provides standard build and installation commands for local inference tools (e.g., cmake, apt install) and model management via CLI (e.g., ollama pull). All commands are contextually appropriate for the stated purpose of AI infrastructure setup.

llm-app-development