Claude Design: A First Look at Anthropic's Move Into the Visual Layer

Anthropic launched Claude Design on April 17, and it took me about an afternoon of poking at it to realize this isn't a Canva clone. It's the visual sibling of Claude Code — a tool aimed at the people who don't have a designer to call but still need to ship something that looks decent. And the most interesting part of it is the bit nobody is talking about: the handoff bundle that drops straight into Claude Code.

The pitch in one sentence: describe what you need, Claude builds a first version, you refine it through conversation or direct edits or custom sliders, and then you either export it (Canva, PDF, PPTX, HTML) or hand it off to Claude Code to actually build. It's powered by Claude Opus 4.7, in research preview for Pro, Max, Team, and Enterprise subscribers.

I've spent two weeks with it across three real projects — a pitch deck, a marketing one-pager, and a mobile app prototype I handed to Claude Code at the end. This post is a working engineer's take on what it does well, where the seams show, and the question I keep coming back to: who is this actually for?

What Claude Design produces

The output formats are broader than I expected. From what I've used and seen:

• Realistic prototypes — clickable, screen-by-screen mockups of mobile or web apps. Real components, real interactions, hover states, the works.
• Product wireframes — lower-fidelity skeletons for early ideation.
• Design explorations — multi-variant comparisons. "Show me five different homepage approaches."
• Pitch decks and slides — full deck structure with consistent type and color.
• One-pagers and marketing collateral — landing-page-shaped artifacts you can ship as HTML.
• Code-powered prototypes — including voice, video, and 3D capabilities. This is the one that surprised me; it's not just static images.

The thing that keeps catching me is the breadth. Most AI design tools are good at one of these. Claude Design is competent at all of them, which makes it more like a generalist design assistant than a specialized prototyping tool.

How the editing model works

This is where the Claude Code DNA shows. You don't just regenerate when something is wrong. You can:

• Leave inline comments on specific elements, like a Figma comment thread that Claude actually responds to.
• Make direct text edits in place. Click, type, done.
• Use custom adjustment sliders that Claude generates per project. For a prototype I built, the sliders Claude offered were "density," "warmth," "playfulness." For the pitch deck, they were different — "formality," "brand intensity." The sliders are not a fixed set of UI controls. They're contextual, generated based on what you're working on.
• Refine through conversation — the standard "make the headline bigger and cooler" loop.
• Group collaboration — multiple people in a doc, leaving comments, suggesting changes.

The combination of inline comments and conversational refinement is the right pattern. I've used tools that only do one or the other and they both feel limiting. Inline comments alone can't restructure. Chat alone can't precisely target an element. Having both means you reach for whichever is faster for the current edit.

The design system feature is the sleeper hit

This is what's going to matter most for teams, and it's getting buried in coverage of the launch.

When you onboard, Claude Design reads your codebase and your design files and builds a design system. Colors, typography, components — the whole token set. From then on, every project Claude Design generates for your team uses your system automatically. You can refine it over time. You can have multiple systems if you have multiple brands.

The reason this matters: the friction in most AI design tools is that the output looks generic. You generate a slide, it's a generic slide. You spend 20 minutes restyling it to match your brand. Then you generate another slide and do it again. The actual unit cost of an AI-generated visual isn't the generation, it's the brand-fitting that follows.

Claude Design eliminates that step for the second slide and every slide after. The first project takes effort because you're getting the system right. Project two through fifty are basically free, because the system is already there and Claude is consistent about applying it.

This is the kind of feature that quietly becomes a moat. Once your team has spent a week getting your design system tuned inside Claude Design, switching to another tool means rebuilding that work. It's the same dynamic that makes a well-tuned CLAUDE.md sticky for Claude Code.

Imports: more than text

The import surface is wider than I expected:

• Text prompts — the obvious one.
• Image and document uploads — DOCX, PPTX, XLSX. You can hand it a sloppy slide deck and ask it to redo the design while keeping the content.
• Direct codebase integration — for the prototype use case, this is huge. Claude Design can read the existing app and design new screens that fit it.
• Web capture tool — point it at any URL, capture an element or section, ask Claude to riff on it. Use this for "here's a competitor's landing page, give me three takes on the same thing for our brand."

I keep using the doc upload one. The amount of design effort that boils down to "make this PowerPoint not look like a PowerPoint" is genuinely high, and Claude Design shortcuts most of it. Not all of it — there's still a pass at the end where you're tweaking specific slides — but the first 70% of the work happens in one shot.

Exports: Canva is the interesting one

Five export options:

• Internal URL — organization-scoped sharing, like a Notion link. The default for "let me show this around before we ship it."
• Canva — full editability inside Canva once exported. This is Anthropic explicitly positioning Claude Design to complement Canva, not replace it.
• PDF — final-form sharing.
• PPTX — for slides specifically.
• Standalone HTML files — for prototypes and one-pagers you want to host or embed.
• Folder — bundle of everything.

The Canva integration is the strategic move. Anthropic isn't pretending to replace the design tool ecosystem. They're saying: we'll do the part you suck at — getting from idea to first draft — and then hand it off to a tool you already use for the polish phase. That's the right call. The work of taking a near-final design and shipping it pixel-perfect is not the work an LLM is well-suited for. The work of getting from blank canvas to something a designer can react to, on the other hand, very much is.

The Claude Code handoff is the real story

Buried in the announcement and barely covered: when a design is ready to build, Claude packages everything into a handoff bundle that can be passed to Claude Code with a single instruction.

I tried this on a mobile app prototype. Built the screens in Claude Design, hit "hand off," got a bundle that included the design system tokens, the screen-by-screen specs, the component library, and a structured prompt for Claude Code. Pasted the prompt into Claude Code, pointed it at a fresh repo. Forty-five minutes later I had a working SwiftUI prototype that matched the design.

Not pixel perfect. Not production. But meaningfully closer than "show Claude Code a screenshot and ask it to build it."

This is the loop Anthropic is building toward, and I think it's the actual thesis of Claude Design, even if the launch coverage is missing it: design and code stop being separate disciplines and become two views of the same artifact. You describe what you want. Claude Design does the visual draft. Claude Code does the implementation. They share a design system and a model. They speak the same language because they're the same model.

That's a meaningfully different stack than Figma → designer → handoff → engineer → code, which is what most teams run today. Whether it actually works at production scale is the open question. For prototyping, it's already working.

Who this is actually for

Anthropic positions Claude Design at "founders and product managers without a design background." That's accurate but undersells it. From two weeks of use, my real-world list:

• Solo founders who need a deck, a one-pager, a landing page, and a prototype, and don't have a designer. This is the most obvious fit. Claude Design replaces about three Fiverr orders.
• PMs who currently make ugly slides. Hand them Claude Design and the slides stop being ugly, which means they get to focus on what's in the slides instead of what they look like.
• Engineers prototyping ideas. This is where I land. I can describe a UI in words; I can't draw it. Claude Design closes that gap and then hands it back to Claude Code.
• Small teams that share one designer across too many projects. The designer becomes the design-system owner and the polish-pass owner. Everyone else generates first drafts in Claude Design instead of stealing the designer's time.

Who it's not for: design-led teams shipping a design-led product. The polish pass at the end is the work, and Claude Design isn't aiming to do that. They've said as much by integrating with Canva instead of trying to replace it.

Where the seams show

Two weeks in, the rough edges:

• Prototype interactions are limited. Static screens look great. Complex interaction patterns — multi-step forms with validation states, animated transitions between screens — are hit or miss. You'll often need to describe the interaction in detail rather than gesture at it.
• The "exploration" mode is uneven. Asking for five variants sometimes produces five interestingly different ideas, and sometimes produces five subtly recolored versions of the same thing. The prompt matters a lot.
• Design system onboarding takes effort. The promise is "Claude reads your codebase and your design files." Reality is "Claude takes a first pass and you spend 30 minutes correcting it before it's right." Worth it, but not zero-effort.
• Enterprise admins, heads up: access is off by default for Enterprise and has to be enabled in Organization settings. Easy to miss, and I watched one team spend an afternoon debugging "why can't I find Claude Design" before they checked.
• Research preview means research preview. Things will change. Don't build a workflow around a feature that might shift in two weeks.

The bigger picture

Claude Design is the first product from Anthropic Labs that isn't a developer tool, and it's shipped under "Labs" for a reason — there's clearly a thesis here that goes beyond "a design tool." Slides, prototypes, marketing collateral, code-powered visuals, voice and video and 3D. That's not a product, that's a category bet.

The category bet is that the visual layer of software is going to be generated, not designed, for most teams who aren't shipping design as their differentiator. The places that are — your Linears, your Lindas, your Notions of the world — will keep designing. The thousands of B2B startups, internal tools, and content teams that just need their work to not look bad will let an LLM do the first 80%.

If that's right, Claude Design is positioned at the front of that wave. If it's wrong, this is an interesting toy that doesn't go anywhere. My read after two weeks is that it's right, but it's right slowly — the first version of this is good enough to be useful for the obvious cases (decks, prototypes, one-pagers) and not yet good enough for the harder ones (full-product redesigns, marketing campaigns at scale).

The right time to try this is now, when expectations are low and the feedback loop with Anthropic is wide open. Build a deck with it. Build a prototype with it. Hand off to Claude Code at the end and see what comes out the other side. If you're an engineer who's been blocked on "but I can't design" — this is the thing that unblocks you.