Vercel built a feedback loop that treats agent instructions like software

Vercel spent more than 200 agent runs building design.md, a public prompt file meant to help coding agents make web pages that feel like Vercel even when they can’t see the company’s internal codebase. The pitch is simple: if a team can name the mistakes that keep happening, it can teach agents to stop making them. The catch is just as simple: that still doesn’t make the agents reliable.

The company laid out the process in a post on Monday. In three desktop scenarios, Codex with GPT-5.5 generated the same six-page task twice — once with design.md loaded and once without it. Vercel’s deterministic checks found 39 known failure modes with the file and 91 without it, a 57% drop in this six-page test. Better, yes. Clean, no. Vercel said every one of the six pages still had at least one failure serious enough to block shipping.

That result is why Vercel moved away from trying to simply copy its internal product design guidance into a public prompt. The company had already used that internal design knowledge for agents working inside its codebase, but once tools live elsewhere, they lose access to that context. A subjective prompt alone didn’t hold up: different models read it differently, and some of the important design and implementation knowledge lived in the codebase anyway. So Vercel built design.md from the ground up instead.

The finished system has three parts. The prompt file tells agents how to make design decisions around copy, hierarchy, typography, color, and publishing, and it also spells out patterns that are off limits. A public stylesheet handles reusable implementation details like spacing and layout. An evaluation loop turns human feedback into updated guidance, while mechanical failures get captured as deterministic checks in code. Vercel also built a local eval app that stores the prompt version, model setup, screenshots, reviewer feedback, and the inputs that produced each page.

Then comes the part that sounds like a normal product workflow, because that’s exactly what Vercel is trying to make it. A separate agent, design-agent, can take a Slack mention, load the current design.md, build the page with the published stylesheet, and post back a screenshot and URL. Weekly feedback from Slack, GitHub reviews, and Figma gets folded together, and repeated complaints turn into proposed changes for human review. The message from the whole exercise is blunt: encode judgment, rerun tests, repeat. Just don’t confuse that with magic.