A Practitioner Tested Claude Against a Real Design System

One design practitioner has loaded Anthropic's Claude with a working design system and recorded what came back. The experiment, published on UX Planet, runs under the headline "Can AI follow a real design system?" and reads as a working diary rather than a benchmark study.

How was the test framed?

The article's subtitle sets the scope in the author's own words: "I Gave Claude Design My Design System and Here's What Happened." The structure mirrors how design leads already evaluate tools in private. Drop a token set and component rules into a model, request a screen, examine the output. That workflow is faster than a formal study and slower than a vendor demo.

It answers one practitioner's real question: will this tool produce something a reviewer can leave in a pull request? The article does not claim to generalize beyond that single desk. That restraint is the strongest part of the framing.

Why does this question matter now?

Design teams have spent recent years consolidating tokens, components, and accessibility rules into versioned systems. Generative AI tools arrived during that same window. Most published evaluations still measure coding performance or copywriting output. Design systems sit between those categories — visual decisions, interaction rules, and accessibility obligations bound into one spec.

A tool that can draft a paragraph but cannot honor a token set does not save a design lead's afternoon. The relevant comparison is whether Claude, or any comparable assistant, can read a system file and produce output a human reviewer leaves alone.

Procurement decisions for generative tools are landing for many design organizations now. Teams need that answer before they sign a renewal.

What does the source actually show?

The available material is a Medium feed item, not the full article. The excerpt confirms one author, one design system, one generative tool, and an unstructured observation phase. It does not surface a scoring rubric, a count of generated screens, deviation rates, or reviewer correction time.

That gap carries information. A single-author, single-system test cannot replace a benchmark study. It can replace the abstract capability claims that vendors ship with their model cards.

How should a team run the same test?

Teams replicating the experiment should fix four variables before they start:

• The exact token set and component inventory passed to the model
• The prompt structure, including any system instructions and few-shot examples
• The first three generations, scored against the design system rule book
• The cumulative correction time a senior designer spends over one work week

Without those four numbers, an experiment is an anecdote. With them, it is procurement-grade evidence that a design director can attach to a tooling decision.

How does WCAG 2.2 fit into the evaluation?

Designers can score design system conformance against WCAG 2.2. Token sets should already map to success criteria such as 1.4.3 (contrast minimum) and 2.4.7 (focus visible). A generative tool whose output satisfies both criteria inside the project's own tokens is a defensible release candidate.

Anything else falls short, regardless of how the pixels look. Teams that cannot pass those two checks against their own library should treat the vendor promise as unproven.

What comes next for these tools?

Anthropic, OpenAI, and Google have each shipped agent surfaces that read a repository and apply project conventions automatically. The next round of design system tests will not measure single prompts; they will measure whether an agent stays inside the system across a multi-file build. Whether generative tools become drafting partners or shipping partners will turn on that single shift.