Vercel Says design.md Cut Known Page Failures 57%, but All Six Pages Still Blocked

Vercel has published design.md, a public guidance file intended to help coding agents produce pages that follow its brand outside Vercel’s own repositories. In the company’s six-page test, the setup reduced previously identified layout and design failures by 57%, but it did not produce a single page ready to ship.

The release addresses a practical gap in agent-assisted design: repository-based agents can inspect the components, examples and design rules beside the code they modify; agents making one-off reports, proposals or pages elsewhere cannot. Vercel built design.md as one URL an agent can load for guidance covering brand, layout, copywriting, responsiveness, design systems and information architecture.

Taste becomes a constrained production system

Vercel says its first attempt simply compressed its internal product-design material into a public prompt. That proved too loose: models interpreted visual language differently, and the prompt lacked the real components and shipped examples available inside a repository. The company replaced it with a three-part system: prose guidance in design.md, a public stylesheet, and an evaluation loop.

The split assigns different work to different layers. design.md tells an agent how to frame the reader’s task, organize evidence and choose a composition. The stylesheet supplies documented classes and design tokens for recurring elements such as headers, tables, stat strips and charts. Because the browser loads that CSS when a page renders, Vercel says the model need not consume the stylesheet code in its context window.

The measurement is about remembered failures

The evaluation process holds prompts, mock inputs and render settings fixed, so changes in output can be tied to changes in the guidance. Vercel uses seven scenarios, from a renewal proposal and benchmark report to an interactive planning page and presentation deck. Full rounds run them on Claude Opus 4.8 and Codex with GPT-5.5; development involved more than 200 runs, according to the company.

Reviewers record corrections against the specific generated page that prompted them. Vercel puts subjective judgment into prose guidance, reusable mechanics into the stylesheet, and mechanically testable problems into code checks. A table repeatedly constrained to prose width, for example, became both an instruction to use available width and a deterministic check intended to catch the error again.

A useful control, not a quality verdict

The 39-versus-91 comparison is narrow by design. It covers only six pages and only failure types Vercel had already observed and encoded as deterministic checks. It therefore measures whether known mistakes recur less often, not whether the pages are broadly well designed or whether the system will be reliable across a larger set of tasks.

Vercel’s next feedback source is real internal use. Its Slack-based @design-agent loads the current file, builds against the published stylesheet and returns a screenshot and deployment URL. The company says repeated comments from Slack, GitHub reviews and Figma are grouped into proposed changes; accepted fixes can alter the agent, design.md, the stylesheet or a check, while entirely new page requests can become future evaluation scenarios.