Semantic Intent

Published: April 21, 2026

Finding a way to achieve a foundational brand consistency in the time of LLMs has been an elusive endeavour for the longest time. AI is incredible at scaffolding applications, but much time is spent supervising outputs, ensuring they don’t hallucinate styles, misalign layouts, and stray from prescribed colour palettes.

Google Stitch announced the DESIGN.md file in March of this year, 2026. This file, which can be read by humans and AI alike, acts as the ultimate ‘system prompt’ for your brand’s visual language. By providing the exact specifications - hex codes, padding, typography, and accessibility rules - an AI can understand your aesthetic and doesn’t have to guess how to style a component.

For me, this comes at an interesting time - I was recently building a similar system for this website, using HTML and CSS. So when I saw this in my feed, it seemed like the natural progression when harnessing the power of AI for design. However, to understand why this is a breakthrough, it helps to look at the ubiquitous nature of the markdown format itself.

The Power of Markdown

Markdown, first created in 2004, has quietly become the backbone of the modern web:

• Documentation: Almost every GitHub repository relies on a README.md to explain how the software works.
• Note-Taking and Knowledge Graphs: Tools that link concepts together (such as Obsidian and Joplin) often rely entirely on markdown files for fast, portable text storage.
• Static Site Generation: Many modern blogs and digital properties (like this one!) compile markdown files directly into web pages.
• Chat Formatting: Most messaging platforms use markdown to quickly format bolding, lists, and code blocks.

Why Markdown for a Design System?

First, it perfectly matches the semantic hierarchy that LLMs natively understand. Because language models are heavily trained on GitHub repositories and documentation, they excel at reading the structured headers, lists, and tables inherent to markdown.

Second, it remains highly human-readable and editable. You don’t need to parse complex JSON objects to tweak a hex code.

Finally, it offers simple integration. There is no need for specialised parsing software or proprietary API keys; it is just a text file you can easily create, share, and drop into the root of any project.

These markdown files act as containers, storing collections of design tokens, each representing a different aspect of the design. Every website, digital property, and app can maintain its own unique version of a DESIGN.md file, holding its specific combination of tokens.

What is a Design Token?

Most simply put, a design token is a semantic name used to store a specific design decision, replacing hard-coded values (like hex codes or pixel measurements) with a clear, reusable label. They strip away the rigid surface data to help systems understand the underlying function (and why).

Modern design systems utilise them because they allow for:

• Systemic Changes: If your brand updates its primary colour, the token’s value changes once, and it cascades effortlessly across the entire app or website.
• Semantic Meaning: A hex code like #FFFFFF just means “white.” But a token named color-background-surface tells a developer - and an AI - exactly where and how that colour should be used.
• Moving Beyond Colour: Tokens aren’t just for palettes; they are used for typography scales, spacing units, border radiuses, and shadow elevations.

Reading Intent

While a traditional variable simply stores a value, a design token is an entire methodology for managing visual decisions across platforms. Crucially, tokens teach LLMs your design intent - the underlying logic and functional purpose behind a component, rather than just its final appearance. By capturing why something looks the way it does, tokens empower both humans and AI to make sweeping visual changes while perfectly preserving the original functional purpose of the design.

Words as Vectors

The most fascinating part of this evolution is how it highlights the multitude of ways we can use language to guide machines. Consider the concept of word embeddings, which maps how words exist within a computer system.

Computers don’t understand language, so they translate every word into mathematical coordinates on a massive, multi-dimensional map. AI groups words by their hidden traits, placing related concepts right next to each other in this spatial matrix. By turning language into geometry, LLMs can literally do mathematics with abstract ideas to predict exactly which concepts should naturally flow next.

Just as word vectors teach a computer the meaning of language rather than just its spelling, design tokens teach a computer the intent of a design rather than just its pixels.

The Agnostic Future: Will this work with other LLMs?

Because DESIGN.md is an open-source, plain-text specification, it is inherently model-agnostic.

You can drop these files directly into AI-assisted IDEs like Windsurf or Cursor. Whether the agent is powered by OpenAI’s models, Anthropic’s Claude, or even local models like Qwen and Llama 3, they can read and execute against these files.

It is worth noting the current landscape is still shifting. Anthropic’s new Claude Design product is currently powered by the Canva Design Engine for visual generation. However, Anthropic’s CLI coding agent, Claude Code, interacts with DESIGN.md flawlessly to write front-end components.

For this format to achieve total industry dominance, traditional design giants like Figma will need to build native “Export to DESIGN.md” capabilities directly into their core platforms.

However, within the AI-assisted coding space, the community is already adopting it as the universal contract between visual identity and AI agents. Time will tell whether the markdown file will fully embed itself into traditional visual workflows, but I am excited to experiment.

- Stav Rose