The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for building agents that ingest untrusted user input, which is an inherent attack surface for indirect prompt injection.
Ingestion points: User messages are ingested via the agent.run() method and the FastAPI /chat endpoint described in SKILL.md.
Boundary markers: The examples rely on the Pydantic AI framework's internal handling and do not explicitly define custom delimiters or "ignore instructions" warnings in the system prompts.
Capability inventory: The templates demonstrate tool calling capabilities (e.g., get_weather, generate_image) and FastAPI integration.
Sanitization: The skill emphasizes Pydantic validation and strict type checking, which enforces data structure but does not necessarily prevent instructional injection within the strings themselves.
[SAFE]: The skill promotes secure coding practices and uses reputable resources.
Secret Management: Instructions explicitly demonstrate using .env files and os.getenv for managing sensitive API keys, avoiding hardcoded credentials.
Trusted Dependencies: The skill utilizes well-known, established libraries including pydantic-ai, httpx, and FastAPI.
Validation: The architecture focuses on type-safe models and automatic validation, which is a security best practice for handling LLM outputs.

agent-builder-pydantic-ai