It writes for a version that no longer exists
The agent used components Primer removed, even though the correct types were installed in the project.
Primer, by GitHub · 10 tasks
- screens didn’t compile
- 8/10
- with rules
- 0/10
Cursor, Claude Code and Copilot ship screens that work, from a prompt or a Figma frame, but skip your components, your tokens and your current version. We measure how often, then install what your AI needs to stop, in two weeks.
Your team already pays for AI. It just hasn’t learned your design system yet.
The same 10 UI tasks given to an AI agent against three popular open-source design systems, with and without preparation.
The agent used components Primer removed, even though the correct types were installed in the project.
Primer, by GitHub · 10 tasks
Fixed margins, sizes and colors instead of design tokens.
Carbon, by IBM · 10 tasks
Neither rules nor a Storybook MCP removed every error. Letting the agent type-check before it hands over did.
EUI, by Elastic · 10 tasks
Read the full reports (Spanish)Independent tests on a small sample: they show trends, not guarantees.
Send us your Storybook or repo. We run 10 everyday UI tasks through an agent and compile every file against your real version. This is a real result.
Everything goes into your repo and your Storybook, so it keeps working after we leave. Your AI stops guessing: it reads your components, follows your patterns and checks its work before it hands anything over.
Your Storybook stories
Instructions for Cursor, Claude Code and Copilot: which components exist, how to use them, and to check their work before handing it over.
Removes: 8 of 10 screens didn’t compilePrimer, before rules
A Storybook MCP, so the agent reads your real props, examples and screen patterns before it writes, instead of guessing from memory.
Each Figma component mapped to the real one in your repo with Code Connect, and tokens kept in sync with Figma variables. When the agent turns a frame into code through Figma’s MCP, it uses your components instead of lookalike divs.
Type checks, design-system lint and UI tests from your stories run on every pull request, for code written by people and by AI.
Removes: 1–2 screens failing, even with MCPEUI: 0 once the agent type-checks
Semantic tokens and closed component APIs, so a hand-typed color or a removed component fails at once instead of in code review.
Removes: 41 hard-coded valuesCarbon, before rules
Full screens as stories, like a sign-up form, a data table or an empty state, so agents compose pages the way your team does.
We run the score again on the same tasks and hand you the report. That number is what our guarantee is based on.
Three products, nothing hidden. Most teams start with the free score and decide after the call.
Step 1
See how often your AI ignores your design system
Free
Report in 2 business days
Step 2
We install what your AI needs, in two weeks
from €6,000
Fixed price agreed before we start, based on the size of your system.
Step 3
Keep it working as you grow
from €1,000/mo
Monthly, cancel after 3 months
Prices exclude VAT. Setup is invoiced 50% on signing and 50% after the final test; Care is billed monthly. Pay by bank transfer, SEPA or card. Need end-to-end tests with Playwright too? Ask us; we quote them with your Setup.
When we finish, we run the same test again. No screen may fail to compile, and your score must close at least half the gap to 100: from 73 to at least 87, for example. If it doesn’t, we keep working at no cost for up to two more weeks.
Still unsure? The free score answers most of these with your own numbers.
An AI-readiness score measures how well AI coding agents such as Cursor, Claude Code and GitHub Copilot use your design system. We give an agent 10 everyday UI tasks, like a sign-up form, a data table or a settings page, with only your package installed, then compile every file against the version you ship. The score, from 0 to 100, combines how much of the UI is built from your components rather than raw HTML, how many screens compile, how many colors and sizes are typed by hand instead of using your tokens, and how many components the agent rebuilt itself. In our tests, GitHub’s Primer scored 79 because 8 of 10 screens used components it had removed, and reached 100 once the agent had rules. A well-typed system can score 100 with no help.
Then we’ll tell you. When types are well documented, the agent reads them and gets it right: in one of our tests it scored 100 with no preparation. The score is free precisely so you can check before talking about anything else.
It’s a good start, but an MCP only carries what’s in your Storybook. In our EUI tests, rules plus MCP reduced errors without removing them; a type-check step took them to zero, in three rounds out of three. That’s why we fix the system itself first: tokens, component APIs, stories and tests.
Yes. More and more teams ask the agent to build a screen straight from a Figma frame through Figma’s MCP server. If Figma doesn’t know which code component each layer is, the agent rebuilds it with divs and hand-typed styles. In Kata Setup we connect your Figma library to your code with Code Connect, keep tokens in sync with Figma variables, and test frames from your own files. Code Connect needs a Figma Organization or Enterprise plan; without it, we put the mapping in your Storybook and your AI’s rules instead.
Not for the score: your design system package or repo is enough. If you don’t have Storybook, setting it up is part of Kata Setup.
Cursor, Claude Code, GitHub Copilot, Codex and any agent that reads project rules or connects over MCP. Rules are tailored to the one you use.
Yes. Angular, Vue or web components take more work, because Storybook’s AI tooling doesn’t document every framework well yet. We scope it in the first week of Kata Setup.
For the score, a public Storybook or a branch is enough. For the rest we work in your repo under NDA, and the tests can run on your own AI account.
Yes. Every piece of this page comes from our own design system, documented in Storybook with accessibility checks in both themes and Playwright end-to-end tests. We also ran our own benchmark on it: 10 tasks based on this site’s screens. It first scored 95 without rules and 98 with them; the gaps it found (no heading or layout components, data types not exported) are fixed, and it now scores 100 in every condition. It covers marketing pages, not app screens such as tables or modals. Read the report or browse the design system.
We’ll send your score within 2 business days. No access to production or data, and no commitment.
Prefer to talk first?
Book a 30-minute call