The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes patterns for ingesting potentially untrusted user content to perform structured data extraction. While this introduces a surface for indirect prompt injection, the skill provides specific defensive guidance.
Ingestion points: Example code in SKILL.md demonstrates processing user-provided strings (e.g., product reviews and invoice text) for extraction.
Boundary markers: The Core Workflow recommends using system prompts to define clear behavioral boundaries and separate instructions from data.
Capability inventory: The skill involves network operations to official LLM provider APIs (OpenAI, Anthropic, Google) for data processing.
Sanitization: The skill mandates the use of Pydantic and Zod for schema validation to ensure the extracted data conforms to expected types before being used in application logic.
[SAFE]: The skill uses and recommends official, well-known libraries and services for its operations.
All dependencies mentioned (such as openai, anthropic, pydantic, and zod) are industry-standard tools for LLM development.
The implementation advice focuses on reliability and security, including model version pinning and avoiding legacy unsafe JSON modes.

llm-structured-output