Claude and ChatGPT: a canvas to bring your artefacts together

First, the honest bit: we don't generate it

Let's be straight about what our tool does and doesn't do. We don't generate components; the models already do that, and better than any wrapper we could bolt on. What we are is the canvas their output lands on, a place to review it, mark it up, map it out and present it.

So this whole post assumes you've already asked Claude or ChatGPT for something and got some HTML back. The interesting question is what happens next, because that is the part the chat window can't help you with.

Paste it straight onto the board

Copy the HTML the model gives you, click onto an empty canvas, and paste (Cmd/Ctrl+V). It drops in as a live component and runs right there, sandboxed: no hosting, no deploy step, no build. A slideshow clicks through, a dashboard's numbers show, a form takes input.

If you'd rather not hunt for the paste, there's an Embed HTML button too. And there are a few 'Made with AI' example boards that show a real component next to the exact prompt that made it, so you can watch the whole loop end to end before doing your own.

Then actually work through it

This is the part the chat can't do. Draw and highlight over the component, drop sticky notes on the bits that need changing, add arrows and callouts. Lay it next to a mind map of your own thinking, so the AI's answer and your questions sit on the same board instead of in two different windows.

And if you want to pull the piece apart rather than just annotate it, you can convert an embedded component into editable shapes and rework it by hand, or rebuild it as your own diagram alongside.

Present it without standing up a website

When it's ready to show someone, full-screen present mode steps through the board with the live components still working, so a mock-up or dashboard demos without you deploying anything first. Or export the whole board as an image to drop into a doc, a deck or an email.

The one real catch: it has to be self-contained

There is one honest constraint. For a component to run on the board it has to be a single self-contained HTML file, all the CSS and JavaScript inline, no external stylesheets, fonts or network calls, because it runs sandboxed. Most AI output already is.

If a piece doesn't render, it's almost always because it reaches out to something external. Asking the model for 'a single self-contained HTML file, everything inline, no external resources' usually fixes it in one line. It's the same constraint that keeps the whole thing private and working offline.