I'm working on a project that needs to dynamically generate simple icons and diagrams. I've tried GPT-4 and Claude - they can output SVG code but the results are hit or miss, especially for anything beyond basic shapes.

Has anyone found a reliable workflow for this? I'm wondering if there are specialized models, better prompting techniques, or if I should just use a traditional graphics library and skip the LLM route entirely. What's actually working in production for you?