design · ai · April 15, 2026
a different take on Claude Design, and it has nothing to do with designers replacement
HTML is the new design canvas
when Google Stitch launched, most of the design community had one reaction: is AI coming for our jobs?
that is the wrong question.
what Stitch was actually showing us had nothing to do with replacement. it was a signal about where the design iteration loop is heading. and when Claude Design launched shortly after, the hypothesis became hard to ignore.
both tools generate HTML. not React. not a full-stack application. clean, well-structured HTML with genuinely good visual output.
that is not a coincidence.
why HTML specifically
when AI tries to generate React code for a full product, it struggles. the output looks like a log. components are inconsistent, libraries conflict, design guidelines get dropped somewhere between the prompt and the render. the complexity of a full application is too large a surface for AI to hold well right now.
HTML is different. it has no logic to wire. no libraries to manage. no state to track. it is pure structure and style. AI can write it fast, write it well, and write many variations of it without breaking.
this is why these tools are converging on HTML. not because it is a limitation. because it is the right tool for the job they are actually trying to do, which is fast visual iteration.
the iteration loop is inverting
here is how design iteration works today. you sketch in Figma, refine across dozens of versions, get to a final direction, hand it to an engineer who translates it into code. the code and the design are never quite the same thing.
here is where this is heading. you prompt for HTML variations. ten directions in minutes. you explore layouts, visual treatments, UX patterns directly in the browser. you pick a direction. then you ask AI to replicate that HTML into a full React application.
the key word is replicate. not create. replication is a fundamentally easier task for AI than creation. when the scope shifts from building something new to faithfully reproducing something that already exists in code, the quality gap closes dramatically. AI reads code better than it reads a Figma file or a natural language description. give it HTML and ask it to turn that into React and the output is more accurate, more consistent, more buildable.
this is the same pattern Replit Agent uses. generate HTML prototypes first. once you like one, build it out. the prototype is not a throwaway, it is the foundation.
the Figr proposal
when I was on contract at Figr, I built a prototype around this exact hypothesis. the idea was to give designers a space to generate multiple HTML variations of a design direction, explore them visually, and once a direction was finalised, push that HTML forward as the basis for the real product. not starting from Figma. starting from code that looks good and iterating there.
prototype: figred-ai.vercel.app
the prototype is not fully functional but the UX of the concept is there. you can see the direction it was pointing.
the proposal did not make it through. but watching Replit, Google Stitch, and Claude Design all converge on a nearly identical pattern in the months after makes the hypothesis feel less like an idea and more like an inevitability.
what this means for designers
if the iteration loop moves into HTML, the design tool of the future is not Figma. it is a browser with an AI that can generate and modify code fast enough to feel like sketching.
the process does not change. the material does.