A while back I started building MermaidCreator, pretty openly a clone of other known apps: a Mermaid editor with a visual canvas, a code editor, and an AI tab that turns a prompt (or a photo of a whiteboard) into a diagram. Next.js, Supabase, Stripe, Claude. Nothing exotic.
The editor itself worked out roughly how you'd expect. The one hard-won lesson there wasn't a feature, it was an invariant: the store is the single source of truth. Visual edits regenerate the code, code edits re-parse into the model, and the preview always renders from the code. Every ugly editor bug I shipped came from updating one side of that loop and not the other.
The thing I actually got obsessed with came later. Architecture diagrams are lies within a few weeks of being drawn. Not because anyone is lazy, the code moves and the picture doesn't. Same for runbooks and onboarding docs. Making diagrams easier to create doesn't help much if they rot at the same speed.
So most of the recent work has been about drift. The app connects to a repo (plus Notion, Confluence, Google Docs), builds an index of docs, Mermaid blocks, API routes, migrations, and dependencies -> no AI in that step, just hashing, and when something a diagram depends on changes, it opens a finding for a human to review. If you accept it, it can push the fix back as a GitHub PR.
One thing I'd recommend to anyone building something like this: point it at itself early. MermaidCreator watches its own repo, and the stats on the homepage come from that. It's found real staleness in my own docs, which was equal parts validating and embarrassing.
I don't know yet whether this becomes a business. But the arc to clone something to learn the space, then follow the actual problem you trip over, has been the most useful part of the whole project.
Top comments (0)