constitutional-ai

Instructor is a structured output library that extracts validated data from LLM responses using Pydantic schemas. It automatically retries failed extractions and provides type-safe JSON parsing with streaming support. Use it when you need reliable, validated data extraction from LLMs like OpenAI or Anthropic.

NeMo Guardrails is a runtime safety framework for LLM applications that adds programmable guardrails. It provides key safety features like jailbreak detection, input/output validation, and hallucination detection using the Colang 2.0 DSL. Use it to enforce safety and compliance rules in production LLM deployments.

LlamaGuard is a specialized 7-8B parameter model from Meta for classifying LLM inputs and outputs across six safety categories like violence and hate speech. It offers 94-95% accuracy and integrates with common deployment tools like vLLM and Hugging Face, as well as NeMo Guardrails. Use this skill to add a robust, dedicated moderation layer to filter unsafe content in your AI applications.

This skill implements Anthropic's Constitutional AI method for training harmless AI models through self-critique and revision. It provides a two-phase approach using supervised learning with AI self-critique followed by RLAIF (Reinforcement Learning from AI Feedback) for safety alignment. Use it to reduce harmful outputs in your Claude applications without requiring human-labeled harmful data.