Thanks, I’ll definitely look into this! I’m planning some benchmark experiments with different prompting techniques since even small model updates (like the 1M token Claude change) seem to change how it uses the tool. If you try it out with Mermaid diagrams, I’d love to hear how it works.