The three steps.
The whole import is three moves, and the third is the one that matters: you see what came across before anything is created.
- 01Open Share in Claude DesignUnder Export, choose Project HTML. Not PNG, not PowerPoint — this reads the structure of the page, not a picture of it.One design is one deck
- 02Download the .zipThe zip keeps images as files, so they arrive as swappable assets. A standalone .html works, but any image it does not carry comes in empty and flagged.Or the .html on its own
- 03Drop it on NewYou get a report before anything is created: what mapped, what did not, and what to do about it — per slide. Then open the editor.Then it is a deck like any other
The well on New. The same well takes a design made in ChatGPT, and a PowerPoint deck.
The prompt to paste, before you start.
A design made with the import in mind lands better than one made without it, so New carries a prompt to paste into Claude Design first. It describes the file the reader wants: the slide canvas, the deck’s twelve-column grid and its margins, your kit’s fonts by name, real text and never a picture of type, data in a real table, one section per slide, and nothing that runs.
It also names every setting the editor can read and what each snaps to — fills, gradients, corners, shadows, borders, grain — so what you set in Claude Design arrives as the same setting here, still editable, rather than as a drawing of it. The prompt is generated from the editor’s own contract, so it is never out of date.
What the prompt asks the design to do
- One section per slide, in the order the slides should run.
- Blocks aligned to the grid; a full bleed is left exactly where you drew it.
- Real, selectable text in the named fonts, nothing under 24px.
- Pictures as image files in the project, one per picture.
- Figures in a real table, so they arrive editable and chartable.
- No scripts, animation, frames or embeds: draw the end state.
What comes across.
The design is measured as a browser would draw it, so what you see in the editor is where things were — not an approximation from the file. Each kind of thing in the design lands as the kind of thing the editor already has.
The mapping the importer applies. Nothing on the left is rasterized; everything in the middle stays editable.
What to look at afterwards.
A design is free geometry; a deck is a grid. Blocks that landed off a column or in the margin are flagged on the slide, and Snap to grid puts them on it in one move. A block that reaches the edge of the slide is taken to be bleeding on purpose and left alone. If the face the design used isn’t one of the deck faces, the style names the one it picked, so you can see what changed before anyone else does.
The same well takes a design made in ChatGPT and a PowerPoint deck.