The topic I replaced Figma with Claude Design, Replit, and Google Stitch for a week — and only… is currently the subject of lively discussion — readers and analysts are keeping a close eye on developments.
This is taking place in a dynamic environment: companies’ decisions and competitors’ reactions can quickly change the picture.
Design tooling has split into two camps over the last year or so. On one side are the manual editors I’ve always lived in – Figma, Affinity, Krita, Photopea, Penpot, etc. On the other are the prompt-first generators, which two years ago I would have written off as a gimmick, but they’ve proven me wrong. There’s serious money behind them now, the tools genuinely seem to understand UI conventions, and the more time I spend with them the less obvious it is where the line between design tool and AI generator actually is.
The question isn’t whether prompt-first tools replace design tools across the board. It’s how much manual control you really need once natural language can handle so much of the structural work. So I put down Figma for a stretch of time in place of Claude Design, Replit, and Google Stitch to see how much they can fill in the work of a “real” design tool.
The framing of Replit as the pure vibe-coding tool is out of date. Design Mode launched in November 2025 as a dedicated visual builder, and in March 2026 it was evolved into the Design Canvas for the Agent 4 release, which is an infinite canvas with live previews and direct manipulation. So Replit is competing for the same designer eyeballs as Claude Design and Stitch now, even if the wider conversation hasn’t quite caught up to that. This was maybe my fifth time using Replit and the first time I noticed direct editing existed, which says something about how buried it is.
The model under it is Gemini 3 (the same family Stitch uses) which means Replit Design isn’t reasoning with anything weaker than the competition, so it understands layout, color, typography, and hierarchy, not just code structure. It took me a minute to actually enter edit mode – there’s a tiny icon at the bottom of the chat bar labeled “select elements”, and that will bring up the editing properties panel. Also noting that you can actually attach a Figma file to the prompt as well, but it doesn’t keep all the editable elements from the file and just interprets it to create functional code.
Here, I was able to adjust font size, weight, alignment, outer space, inner space, color, opacity, and the border. Basically what I’d expect from a design tool, not a code generator. Simple changes apply directly and don’t consume credits, only structural rewrites do, which was a nice surprise. In practice, it covers more of my Figma workflow than I expected. Testing type adjustments, swapping color systems, and getting a screen to a presentable state overall can all happen with Replit’s edit controls. But beyond text, layout increments, and colors, you can’t do much else, including dragging, so you have to reprompt to make significant changes to the larger layout.
Claude Design launched April 2026 under Anthropic Labs, runs on Opus 4.7, and you can find it at claude.ai/design. It’s on Pro and above, which means free users are still locked out unfortuntely. But you’re getting what you pay for from Opus 4.7 which is currently the strongest reasoning model sitting behind any of these tools and you can feel it in how it interprets ambiguous prompts. When I say “calm” or “playful” it actually translates that into structural decisions regarding spacing, hierarchy, weight, and not just a color shift.
File uploads aren’t unique to Claude here (Replit and Stitch both accept image uploads), but Claude’s are the most flexible. DOCX, PPTX, XLSX, images, and direct Figma file imports all integrate cleanly, so if I’m coming in with existing material from anywhere else, this is the easiest landing pad. And it achieves higher fidelity to the original Figma design than Replit does.
The real differentiator is what happens after Claude generates something. The Tweaks panel spins up contextual sliders specific to whatever was just made, which can be color palettes, font weights, spacing, padding, corner radius, whatever Claude thinks is worth adjusting (but you can also request them yourself). Nothing else in this trio does this on the same level. I still can’t drag things around (layout moves still need a prompt), but for fine-tuning everything else, I’m not burning tokens on adjustments I could just make with a slider, and actually get the most relevant adjustments handed to me based on the context.
The tradeoff is that even on Pro the credit ceiling chews through faster than I’d like. But the workflow feel is closer to “designer space with an AI inside it” than “AI tool with some buttons”.
Stitch sits inside Google Labs, runs on a Google account, and is free (you can find it at stitch.withgoogle). It launched at Google I/O 2025 and got rebuilt in March 2026 into an AI-native canvas with multi-screen generation, infinite canvas, voice input, and prototyping. Google calls the approach vibe design, where you describe the feeling and it handles the structure. Standard mode runs Gemini 3 Flash and Experimental mode runs Gemini 3.1 Pro.
Compared to Claude Design being paid-only and Replit’s daily Agent credit ceiling, Stitch is by far the most accessible of the three. Manual control is distributed differently than Claude’s: there are fewer contextual sliders, but the palette panel is actually a full design system editor where you can tweak light/dark toggle, seed color, primary/secondary/tertiary/neutral swatches, font picker, and corner radius. And adjusting these doesn’t require using your prompt limits. So while it doesn’t have as much microscopic control as Claude and Replit, it does consider the design as whole and lets you adjust it on the system level, basically altering this palette applies it to everything on the screen.
There is one tool that gives you a little more fine-grain control though, and that’s the Direct Edit pencil icon. It doesn’t bring up individual properties like spacing or color, but it does let you click on a single element and adjust it using the AI. This means I don’t have to reprompt the entire design, and instead make element-level adjustments using natural language, which is actually one of the best behaviors a vibe-designing tool could have. It doesn’t waste tokens on redoing the whole thing.
To me, Claude Design takes the crown, and it’s because of the Tweaks panel. These are manual controls just like you’d get in a traditional design tool, but they’re created on the fly and context-aware. That’s probably one of the smartest features I’ve ever seen in a creative tool. Beyond just being impressive, they’re actually incredibly useful. How many times have you made manual adjustments in a design tool when something feels a little off, going through all the controls from top to bottom just to see what actually works? Claude solves that with Tweaks and gives you the exact thing you actually need to adjust.
Google’s Stitch is a close second though. The combination of the system-wide palette plus the individualized prompt box gives you full control from the system level all the way down to single components. I don’t want to write off Replit though, I just feel like its design mode is a little undercooked compared to the other two. And this is not a big surprise given its roots is in vibe-coding not vibe-designing.