Digital Product Talks #40: This is why AI keeps ignoring your design system



The story about design and AI is often the same. You open Figma Make, or Lovable, or Claude, ask for a screen, and a few minutes later you have something. It’s like ✨ magic. ✨
However, though it looks fine, it uses none of your components. Let’s not even start on what it has invented.
The reason is boring. Your design system exists as a drawing, not as writing. AI reads names, values and tokens fine. It can’t do what a new designer does: open the file, hover around, and work out the why.
Andreas Kleinadel, Head of Digital Product Design at Handelsblatt, put it more bluntly when he came on Digital Product Talks.
“Whenever we experiment and say, dear AI, let's build design XY, or I need this or that product, it usually goes wrong. And I think it goes wrong because the AI doesn't understand how our design system works.”
You can fight it in chat all you want, but what a design system needs is a source it can read and understand. By only telling it what you need, it’s working off assumptions, so it starts inventing things like it’s insecure you’ll judge it if it says it doesn’t know what to do.
Andreas has five designers (six including him) covering two 80- and 100-year-old brands, and a multi-brand, design-to-code design system they've been building for two and a half years. By most measures, that system is in good shape.
Here's where his team is putting most of its time now:
“Our design system documentation is so much more important than building the next component, and the next component.”
Documentation, and precisely because of AI, as that’s a format it can understand (say thanks, AI). Unfortunately, no designer has ever said they love doing documentation.
Let’s hear why documentation is so important for design systems and AI.
.gif)
How does a new designer learn a system? They open the file, look around, hover around, study the variants, copy, test, try. If something is unclear, they ask someone who can show them. Almost none of it is written down anywhere in words.
AI can’t learn that way. It understands names and values, and tokens that give it your colours and spacing. But it doesn’t understand why. When do you use a secondary button? Why is a destructive dialogue different from every other dialogue? Why does an error state look the way it does?
So it fills in the gaps on its own, based on the average of every design system on the internet. So that’s what you get in the output: a generic dashboard, or as the kids say, AI slop.
Andreas described the fix as a translation job:
“We have to translate our UI, our experience principles and what makes up our brand into language AI can understand.”
This is uncomfortable for a design team, and not because they prefer rectangles to writing, but because… well, actually, yes. Translation is writing, and writing documentation is, unfortunately, not fun.
When making your design system readable, here is what you have to keep in mind:
There’s also the technical part. Handelsblatt is building an MCP server so their design system can be read by AI tools directly, and the finding from that work should be your #1 rule:
“The better the elements and the content are described, the better the result that comes out.”
MCP is how the system gets reached. Documentation is what's there when it arrives. A badly documented system served over MCP is faster access to a proper mess.
.gif)
Halfway through the episode, Andreas turned the question around on us and asked what COBE’s experience is. Fair!
Your seniors probably won’t like step three, but that’s until the first prompt comes back using your own components, with your own rules, reducing the review time to four minutes instead of forty. Everyone is happy then. 😎
The original promise of the design system was that the repetitive part would get cheaper and the thinking part would get more room. However, after spending a few years getting very systematic, somewhere along the way people stopped talking about experience. Everyone talks about components now.
Nothing is wrong with your design system. It just can't be read yet, and the fix for that is in your Figma file right now, just undocumented. Getting it out takes a few very boring weeks of writing, but when AI takes the assembly work, the room for thinking comes back.
If you're writing documentation for machines and want to compare notes, we're around.
Want to hear the full conversation with Andreas Kleinadel? Listen to the episode on the Digital Product Talks podcast on Spotify or Apple Podcasts.
Why does AI ignore my design system? Because it can’t read it. A design system lives in Figma as components and variants, which AI can’t interpret the way a designer can. It reads names, values and tokens, so everything else gets filled in from the average of every design system on the internet.
What makes a design system AI-readable? Written documentation covering four things: component states (loading, empty, error, disabled), the “never” rules that say what a component must not be used for, content rules for labels and error messages, and experience principles specific enough that they couldn’t describe any other product.
Does an MCP server fix it? No. A Model Context Protocol server is how AI tools reach your design system directly. But it only serves what’s there. A badly documented system served over MCP is faster access to a mess.
Felix is the CEO and one of the co-founders of COBE always looking for ways to make the world a little more beautiful.




