Google Labs' Stitch, the AI-powered design tool, is open-sourcing the draft specification for DESIGN.md, a portable markdown file format that lets teams export and import design rules between projects so brand guidelines travel with a codebase rather than being rebuilt each time.
DESIGN.md encodes the reasoning behind a design system, capturing colour palettes, typography hierarchies, spacing systems, component patterns and interaction specifications in a format that is both human-readable and machine-readable.
When paired with Stitch's design agent, which is powered by Gemini, the file functions as persistent context that shapes every generated interface, ensuring output matches a team's brand without requiring design rules to be specified in each prompt.
"Instead of guessing intent, AI agents can know exactly what a colour is for, and can validate their choices against WCAG accessibility rules," the company said.
The format follows a similar logic to agents.md and rules.md files in the developer world, setting explicit instructions for AI using a file that both humans and machines can read and write.
Without such a file, AI-generated interfaces tend to produce inconsistent results, with different button styles, spacing and colour choices across screens.
With DESIGN.md, the model applies the same visual rules throughout a project.
The draft specification is being released as open source on GitHub to allow the format to be adopted across tools and platforms beyond Stitch.
Through Stitch's MCP server and software development kit, DESIGN.md can connect to development tools including Claude Code, Cursor and Google's own Antigravity, meaning visual specifications defined by designers can be automatically followed when developers write code.
Users can create DESIGN.md files in three ways: by extracting design systems automatically from any URL, by uploading brand assets such as logos or visual identity manuals for AI analysis, or by building a file from scratch.
The format uses markdown rather than JSON or YAML, a deliberate choice because large language models parse markdown more reliably and the format allows for richer contextual information alongside raw design tokens.
Related reading
- OpenAI closes $122bn funding round at $852bn valuation as IPO pressure builds
- Google makes Gemini Embedding 2 generally available as its first natively multimodal embedding model
- Trump signals thaw in Anthropic standoff, says Pentagon deal 'possible'
Google Labs' David East explains the format and its intended use in a companion video released alongside the announcement.
The community has already begun building around the specification, with an open-source repository of ready-made DESIGN.md files extracted from the public-facing websites of companies including Figma, Vercel and Revolut now available on GitHub.
The recap
- Stitch publishes draft specification for the DESIGN.md format.
- AI agents can validate choices against WCAG accessibility rules.
- Users can generate files in Stitch or contribute on GitHub.
