← Tous les articles

AI Tools · September 9, 2026 · 4 min de lecture

AI Can Design a Beautiful Screen in Minutes. It Still Can't Build Your Design System.

AI tools like Claude and Lovable can produce a polished screen in minutes. Without a design system behind it, every new screen quietly drifts from the last one.

Someone on your team opens Lovable or Claude on a Friday afternoon, describes the product idea in two sentences, and by four o'clock there is a working screen: a clean landing page, or a dashboard with real components, animations, the works. Everyone in the room is stunned. This is the fastest anyone has ever gone from idea to something you can actually click on.

The following week, someone asks for a second screen, a settings page, or a pricing table. It comes back fast again, and it looks fine on its own, but next to the first screen the button corners are a different radius, the spacing is looser, the blue is not quite the same blue. Nobody can point to what went wrong, because there was never anything to go wrong against. There was no system, only a very good first guess.

What's actually happening

Tools built for this moment are genuinely capable now, not just impressive in a demo. Lovable's Agent Mode can explore a codebase and fix things on its own, its Visual Edits give a Figma like click to adjust experience, and its Draw to Build feature turns a rough sketch into working code. Claude Design, which Anthropic launched in April 2026, goes further: it can build a design system from scratch or work from one you already have, produce wireframe flows and high fidelity prototypes, hand working frontend code to engineers, and read Figma files directly through the Dev Mode MCP server so the design and the code stay in sync.

None of that is the problem. The problem is that most teams are still using these tools the way they used a single Figma file: one prompt, one screen, judged on whether it looks good in isolation. AI does not drift on purpose, it drifts because every prompt is answered fresh, with no memory of the constraint that made yesterday's screen consistent with today's, unless someone deliberately gives it one.

The reframe

Most people conclude that if AI can produce a polished screen in minutes, the formal design system, the color tokens, the spacing scale, the documented component names, matters less than it used to. What actually matters is closer to the opposite. The faster and cheaper it becomes to generate a screen, the more that screen needs a system to be generated against, because now the drift compounds at AI speed instead of designer speed. A team that used to produce three inconsistent screens a month can now produce three a day.

This is where Alignment, one of the later questions in a real AI adoption process, shows up before anyone expected it to. A design system is not decoration for a mature company. It is the constraint that keeps a fast, confident tool coherent across pages, flows, and the five different people now prompting it.

What this looks like in practice

  1. Give the AI tokens, not adjectives. Define your color palette, spacing scale, type scale, and named components once, in a short written brief, and paste it in before the first prompt of a session. "Modern and clean" produces something different every time. A token list produces the same visual language every time.
  2. Treat every AI generated screen as a draft against the system, not a finished decision. Someone still needs to decide whether a new pattern earns a place in the system or gets pulled back in line with what already exists.
  3. Save the brief once, reuse it every time. Claude projects and Lovable both let you attach persistent context to a workspace. Paste the design system in as project knowledge so every new session starts from the same constraints instead of reinventing them.
  4. Name one person who owns the system. Someone has to notice drift and decide, deliberately, whether to update the system or correct the output. Left undecided, the newest AI generated screen quietly becomes the house style by accident.

Speed was never really the bottleneck in product design. Coherence across screens, teams, and time was. AI did not remove that bottleneck, it just moved it earlier, into the fifteen minutes before the first prompt instead of the fifteen hours after the tenth screen shipped.

Try this week

Open a blank document and write down five things: your core colors with hex codes, your spacing scale, your type scale, your corner radius, and the names of your three most used components. Paste that list as the very first message the next time you open Claude or Lovable, before you ask for a single screen.

Sources

  • Best AI App Builders in 2026: Top 6 Tools Compared
  • 8 Best UX Tools for 2026: Design, Test & Build Apps
  • Claude for UX Designers: The Complete 2026 Guide to AI-Assisted Design Workflows
  • 10 Claude Skills for Design To Improve Workflows
  • The State of AI in UX & Product Design: 2026
Partager cet articleLinkedInFil RSS