I'm having the LLM work with human-readable behaviors for each component. You can have the LLM take those behaviors and create a Mermaid diagram, for instance.