Anthropic's design tool: type a prompt, get a live interactive HTML page in the canvas. Made for people who do not write code.
You can describe what a site, deck, or prototype should look like, but you cannot draw it from scratch and you do not write code. Claude Design closes that gap: you talk in the left pane, a live canvas renders in the right pane, and you refine through chat, inline comments, or direct canvas edits. The output can become a PDF, PPTX, standalone HTML, connector handoff, or Claude Code build path depending on what you need next.
Steps
Go to claude.ai/design, or open it from the Claude Desktop sidebar when your app shows it. Claude Design is in beta for Pro, Max, Team, and Enterprise plans; Enterprise accounts usually need an admin to enable it first. You'll see a chat pane on the left and a canvas on the right. Type a real description: who the site is for, the feeling, two or three reference brands, the sections you want. The richer the first prompt, the closer the first render is to where you want to end up.
The canvas is live. Use chat for broad changes like 'replace the case study cards with a horizontal scroll.' Use inline comments for one element. Use direct canvas controls for quick text, drag, resize, and align fixes. You are not editing source files by hand; you are choosing the steering move that fits the change.
Claude Design can bring in one or several design systems from a GitHub repo, design files, raw uploads, or a local codebase through `/design-sync` in Claude Code. You can also provide screenshots, decks, docs, logos, color palettes, typography notes, or a DESIGN.md file. The cleaner the source material, the better the generated system and the less Claude has to invent.
When the canvas matches your intent, choose the destination by the final proof. Export PDF for review, PPTX for a deck, standalone HTML for a static file, Send to Canva or another connector when that tool owns final editing, or hand off to Claude Code when the design needs to become a working site. Inspect the result after it lands; the destination owns the final QA.
How do you know it worked?
If something looks off
See also