Deep dive: M3E Canvas turns a real M3 component canvas into a 4,000-char prompt

[LAIguapi]

I spent an evening with M3E Canvas and wrote up a hands-on deep dive (in Chinese): pulled the code, read through the architecture, and built a 3-screen expense-tracker prototype to see what the generated prompt actually looks like.

A few things that impressed me:

• Zero LLM inside. The canvas-to-prompt compiler (lib/prompt.ts) is fully hand-written in three languages. No API key, deterministic output, works offline.
• The prompt quality is remarkable. My 3-screen sketch compiled into a 4,066-character Chinese spec: full color-role hex values, per-screen layout, tap behaviors (including "reverse the entering transition on back"), and M3 component specs. My free-text notes on components were correctly woven into the behavior section.
• The magnetic snap/fuse interaction is genuinely fun — drag chips together and they fuse into a connected group with correct M3 Expressive corner rounding, and the prompt describes it as a horizontal chip group automatically.

Friction points I hit: default language is Japanese (browser locale saves you), default light theme gets baked into the prompt ("light mode only" — easy to miss), and it locks to preview-only mode below 1200px width.

Great work — this fills a real gap between freeform whiteboards (tldraw/Excalidraw) and black-box generators (v0). The readable, editable prompt as an intermediate layer is an underrated design choice.

[lnkiai]

Thank you so much for the deep dive and the kind words! If the Chinese write-up is published somewhere, I'd love a link.