What AI design tools still leave to designers

Two releases this spring changed what the first week of a product looks like.

On 18 March, Google relaunched Stitch, its free design experiment from Google Labs, around what it calls vibe design. You describe the business goal, the feeling you want or a few reference sites, and a design agent produces screens on an infinite canvas. Press play and the screens become a clickable prototype. Then, on 17 April, Anthropic released Claude Design, a research preview for paid Claude plans that turns a conversation into prototypes, slide decks, one-pagers and landing pages, and can hand the result to Claude Code to build.

They are fast, and for some jobs they are good. They also have a clear limit, and knowing where it sits will save you money.

Where they help

The obvious win is speed at the start, when the expensive mistake is building the wrong thing.

• Exploring directions. Ten layouts for a home screen in the time it used to take to sketch one, so you can react to real options instead of describing them.
• Clickable prototypes for early users. Watching five potential customers try to book an appointment in a prototype teaches you more than any meeting about requirements.
• Investor and sales material. A pitch deck or a one-pager that looks considered, without a week of back and forth.
• Landing pages to test demand before the product exists.
• Giving developers a concrete target. A prototype says more than a written spec, and with Claude Design the handoff to code is built in.

Where they fall short

The weaknesses show up later, when a prototype has to become a product people rely on.

The first is sameness. These tools learn from what already exists, so left alone they produce competent, familiar layouts: the same hero section, the same cards, the same rounded buttons. Fine for a test. A weak foundation for a brand that needs to stand out.

The second is the hard thinking. An AI can draw a screen for any flow you describe. It cannot tell you which flow your customers need, what to leave out, or what happens when the payment fails, the internet drops or the user is on a five-year-old phone. Most of a designer's value was always in those decisions.

The third is everything beyond the happy path: empty states, error messages, loading, accessibility for people using screen readers or larger text, and right-to-left layouts for Arabic. Generated designs often skip these, and Arabic layouts in particular need checking by someone who reads Arabic.

Why your design system matters more now

Both tools can work from a set of design rules instead of starting fresh. Claude Design reads your codebase and design files during setup to build a design system for your team, then uses your colours, typography and components in later projects. Stitch introduced DESIGN.md, a plain text file describing your design rules that other AI tools, including coding assistants, can read.

It is the same idea as Agent Skills or llms.txt: write down how you do things in a form machines can follow. Here it means that a business with a clear, written design system gets designs that look like its own. A business without one gets designs that look like everyone else's.

So if you plan to use these tools seriously, the most valuable design work you can pay for is the system itself. That means colours, type, spacing, components, tone and the rules for using them, documented properly and kept in sync with your code.

How we would use them

For a new product, we would use AI design tools in the first week to explore and test quickly, with real users clicking real prototypes. Once the direction is clear, a designer sets the system and works through the hard flows and edge cases. After that, the AI tools become useful again, because they can now produce new screens that follow your rules instead of guessing.

The tools will keep improving. The questions about who your customers are, what they need from you and what you want them to feel will still be yours to answer.