Correct me if I'm wrong: Anthropic is a startup working to develop guardrails for real AI concerns.

In general, how does one put "business rule"-type constraints on the output of models? Is there any way to bake-in prime directives?