If you're tired of writing brittle JSON parsing code and handling LLM output errors manually, Instructor wraps Anthropic and OpenAI APIs to enforce Pydantic schemas on responses. Define your data structure once, and it automatically validates outputs, retries with error feedback when validation fails, and gives you proper Python objects instead of raw JSON strings. The streaming support is solid for real-time UIs, letting you show partial results as they generate. Over 15,000 GitHub stars and used by 100,000+ developers, so the edge cases are probably handled. It's the difference between hoping your LLM returns valid JSON and knowing it will.
npx skills add https://github.com/orchestra-research/ai-research-skills --skill instructor