Constitutional Ai

This teaches Claude how to apply Constitutional AI principles to its own outputs. It's based on Anthropic's technique where models self-critique and revise responses using a predefined set of principles, basically getting the AI to police itself for harmful content without needing human labels. The skill covers both the supervised learning phase where Claude learns to critique its own work and the reinforcement learning phase using AI feedback. With 325 installs and solid audit scores, it's worth trying if you want Claude to be more thoughtful about safety guardrails in your project. The approach is clever but you'll want to define clear principles upfront or the self-critique won't be useful.