Claude Design and reality: what AI is actually good for in design work

Anthropic launched Claude Design on April 17, 2026: you describe what you need (a landing page, an app, a pitch deck, a marketing one-pager) and the tool delivers a first version within seconds, as presentation-ready HTML/CSS/JS. VentureBeat and several industry blogs described it at launch as a challenger to Figma. Refinement runs through four channels: chat, inline comments on specific elements, direct text editing, and sliders Claude itself generates to adjust spacing, color and layout live.

So this isn't an "AI draws me a website" gimmick, it's a genuine iteration tool. But there are two things the demo videos rarely say out loud, and we hit both of them in our day-to-day work.

What it's genuinely good at

• Recognizing and following an existing visual system. During onboarding, Claude Design reads a team's codebase and design files and builds a design system from it (colors, typography, components), which it then applies automatically to every subsequent project. We enforce the same principle by hand on our own landing page: a single accent color (a cool blue glow) reserved strictly for CTAs and the logo; everything else stays on a neutral, dark base, because an undisciplined palette competes with the call to action.
• Deterministic, data-driven components. Our own wheel diagram (a static SVG donut with a logo hub in the center) is fully props-driven: you can add any number of modules with zero code, because the visual logic lives in a component, not hardcoded SVG.

What the demo videos don't show you

Anthropic itself calls Claude Design a research preview, and the limits are concrete: no backend generation, it's a purely front-end, visual output tool. For brand-critical, conversion-sensitive interfaces, it's a starting point, not a finished product: early reviews make the same point.

Our own experience confirms this on three specific fronts:

• Never recreate brand logos "from memory." For one technology partner's logo (GoHighLevel), the model once tried to recreate it as inline SVG from its own knowledge, and the result was unrecognizable. The fix: always source the real vector from an official source (web search, favicon zoom-in), never rely on the model's visual memory.
• Mobile layout can only be fixed from a real screenshot. A vague complaint ("the button is too big on phone") easily gets fixed in the wrong place based on a description alone; without a screenshot, the fix is a guess.
• iOS Safari is the toughest test environment. A glow effect that disappeared perfectly in Chrome at the end of an animation got stuck on iPhone; adding visibility: hidden at the animation's closing frame fixed it. This only shows up on a real device, never from assumption.

Our principle

Claude Design (and Claude in our design work generally) doesn't replace design decisions, it speeds up the execution of good ones. The design system (color, rhythm, component logic) is always built with human sign-off: the model writes the code fast and consistently, but on the question of "what should look good," we decide, and we verify every visual change on a real browser and a real device before shipping.

Sources: Anthropic: Claude for Creative Work, VentureBeat: Anthropic just launched Claude Design