jiangdayuan

I’m working on an open-source thing that lets you co-draw diagrams with an LLM inside draw.io/diagrams.net.

You can ask the model to rough out an AWS/GCP/Azure architecture, but the key part is the loop: you still have the normal editor, so you drag boxes, rename stuff, add your own bits, and then say things like “clean this up”, “split this service out”, “add a read replica here”. The AI edits the real draw.io XML, it’s not just generating a picture, so you and the model are basically sharing the same canvas.

It can also try to rebuild a diagram from a screenshot/PDF and then you keep iterating together in chat + manual edits.

Recently I added “bring your own API key” for a bunch of providers and support for uploading PDFs/text to turn existing docs into diagrams.

Repo (just crossed ~10.2k): https://github.com/DayuanJiang/next-ai-draw-io Demo: https://next-ai-drawio.jiang.jp/

If you live in drawio a lot, I’d be curious where this breaks down or feels more annoying than just doing it by hand.

Thanks! I hope it helps with your workflow. Let me know if you find any bugs or have feature requests!

I think you're misreading the comment. The reasoning isn't missing, it's implied by decades of good design principles: simpler is better. A normal developer would intuitively prefer <output>, so the question isn't "Why suggest <output>?" but rather "Why is <output-format> necessary?"

And in the context of LLMs, this isn't just a matter of aesthetics. More verbose tags mean more tokens, and more tokens mean higher costs. It's a perfectly valid and practical piece of feedback.

The performance of Gemma 3 is insane.