The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for building llama.cpp from source which involves using sudo for system-wide installation and package management (e.g., sudo apt install, sudo cmake --install in references/local-inference.md). This is consistent with the skill's purpose of setting up local inference environments.
[EXTERNAL_DOWNLOADS]: Installation instructions for Ollama in references/local-inference.md involve downloading a shell script from ollama.com (curl -fsSL https://ollama.com/install.sh). Ollama is a well-known service for local LLM inference.
[REMOTE_CODE_EXECUTION]: The skill documents the installation of Ollama by piping a remote script to a shell (curl -fsSL https://ollama.com/install.sh | sh in references/local-inference.md).
[INDIRECT_PROMPT_INJECTION]: The skill documents architectures for RAG pipelines and AI agents that ingest untrusted data from external sources (e.g., documents, tool outputs).
Ingestion points: Untrusted data enters the context through search results in RAG examples (SKILL.md) and execute_tool results in agent loops (references/agent-patterns.md).
Boundary markers: The skill explicitly recommends and demonstrates the use of XML tags and delimiters (e.g., <user_query>, <retrieved_documents> in references/safety.md) to isolate data from instructions.
Capability inventory: Code examples utilize network requests to LLM APIs, database interactions via psycopg, and arbitrary tool execution (execute_tool in references/agent-patterns.md).
Sanitization: The skill includes a detailed safety reference (references/safety.md) that covers PII stripping using Presidio, output validation with Pydantic, and prompt injection defense patterns.

llm-app-development