Google open-sources Stitch's DESIGN.md to teach AI your design rules

DESIGN.md is an open format specification designed to give coding agents a structured, persistent understanding of your visual identity. Originally developed for Google Labs' Stitch UI tool, the draft specification is now completely open-source on GitHub.

Basically, instead of pasting random color codes into system prompts and hoping your AI remembers them, DESIGN.md packages your entire design system into a single portable file that any agent can read and follow.

What it does

Anyone using AI to build interfaces knows the pain of styling drift. Models hallucinate border radiuses, invent new hex codes, and forget brand hierarchy after a few prompts.

DESIGN.md fixes this by establishing a standardized design contract. You can export rules from Stitch and import them into any platform or agent workflow.

Crucially, it teaches agents the reasoning behind design choices while enforcing WCAG accessibility rules. Instead of guessing whether white text looks readable on a bright primary button, the agent can actively validate contrast ratios against official accessibility standards.

How it works

Google solved this without inventing a weird proprietary format. A DESIGN.md file simply fuses two complementary layers: YAML front matter for machines, and Markdown prose for intent.

The YAML layer defines explicit design tokens. This includes your palette, typography scales, spacing, border radiuses, and base component variants. These provide the hard, normative values agents need.

The Markdown body provides the design rationale. Structured into ordered sections like Overview, Colors, Typography, and Do's and Don'ts, it tells the agent why those values exist and how to apply them. For example, if you define an accent color called Boston Clay, the prose clarifies that it should serve as the sole driver for interactive calls to action.

Google also shipped practical tooling to go with it. Using simple npx commands, developers can lint a file to surface contrast warnings and broken token references as structured JSON, or diff two versions to track regressions across iterations.

How to try it

The draft spec is available on GitHub under the google-labs-code repository, currently at the alpha stage.

If you already use Stitch, you can export and import DESIGN.md files directly inside the app. For other coding agents, you can grab the spec documentation from GitHub, drop a template into your repo, and use the official linting package via npm to validate your design tokens.

Why it matters

If you use AI coding agents to build frontend interfaces, this is worth testing immediately. It replaces bloated, fragile style prompts with a clean format that validates accessibility out of the box.

We think open-sourcing this format is a smart play. Google could have easily locked DESIGN.md inside Stitch to build a proprietary moat. By making it an open standard, they are trying to solve the missing bridge between human design tokens and AI code generators across the entire developer ecosystem.