Setting up a design file for your coding agent is supposed to take almost no effort. Better Stack’s guide by Stanley Ulili says it is enough to put the file in the project root and tell the agent to “strictly follow DESIGN.md”. It describes no verification step and no linter run. The file in question is DESIGN.md , Google Stitch’s format for describing a visual identity to coding agents. Here is the problem: a file meant to make AI-built interfaces follow the design system often tells the AI to use text colours many people can’t read, and the official checker lets it through.
What the first large dataset measured
The format’s README is blunt about authority. It says “The tokens are the normative values”, and it describes the file as giving agents “a persistent, structured understanding of a design system.” Its linter has a contrast-ratio rule for component background and text pairs below the WCAG AA minimum of 4.5:1. That rule is a warning. Only a broken reference counts as an error, and the CLI exits with code 1 only “if errors are found”, 0 otherwise.
On 13 November 2026, Jialiang Wei and Walid Maalej published DesignSpecs , a dataset built from 613,207 files named DESIGN.md collected in October. Of those, 36,906 were visual design specs that passed the linter with zero errors. Among the 7,817 unique contents that define component tokens, 43.2% have at least one text and background pair below 4.5:1. Another 88.8% carry colour tokens no component references.
So “passes the linter” and “meets the contrast floor” are different bars. The paper itself notes that passing “does not certify design quality”.
The decision moved before the interface exists
Ask what this means for a front-end developer or design engineer who drops a DESIGN.md into a repo, whether their own, Stitch-generated or copied. They are handing the agent colour pairs it will apply to every button and card it builds. That makes the file the place to check contrast, before any UI exists.
Copying sharpens the point, a theme we met in the forkable brand files piece. According to the dataset, copies beyond the first instance of each unique content account for 19,666 files, or 53.3%. The most-shared file, “Apple-design-analysis”, appears as byte-identical copies in 208 repositories. One file’s colour decisions become the “normative values” for 208 projects’ agents. The paper does not say whether that file carries contrast warnings, so it shows how decisions spread, not a confirmed failure.
The accessibility decision has moved upstream into a small spec file, so the audits teams already run examine the product after the choice has been made.
Here the inference needs marking. DesignSpecs measured file contents, one commit per repository, plus linter warnings on token values. Google’s README documents what the linter flags and how it exits. Neither measured rendered interfaces or agent behaviour. Read across, though, the two together suggest that agents handed a failing pair will probably build it faithfully, and that copying will carry it further. That is a step from “the spec says this” to “the product will probably do this”, and the step is ours.
It also explains why the usual safety net may miss it. The accessibility decision has moved upstream into a small spec file, so the audits teams already run examine the product after the choice has been made. Accessibility checks, like the automated scans we looked at in September , run against the generated UI. By then the colour was a requirement, not a defect.
What the warning might still be right about
The strongest objection comes from the paper itself. Its contrast check examines token values, not rendered interfaces. The linter compares every component pair against 4.5:1, while WCAG allows 3:1 for large text, so some flagged pairs may be compliant in use. Agents may also ignore or override tokens. And a warning rather than an error can be defended as leaving room for deliberate design choices, though only a third-party explainer, instagit.com, makes that case, not Google.
All of that is fair, and it cuts the 43.2% from a defect rate to a risk indicator. It does not remove the risk. No source measures whether agents given these files actually ship low-contrast UI, so our claim stays a probability. But the baseline is not reassuring. The WebAIM Million found low-contrast text on 83.9% of home pages in 2026, up from 79.1% in 2025, and names automated or AI-assisted coding practices as likely contributors, not proven causes.
The practical response is cheap. Run the linter, but read its warnings as the first accessibility review rather than as noise. A file that validates has only proved it is well formed. Whether the pairs inside are legible to a person is a separate question, and today nothing in the format asks it loudly enough.



