Plaintext Taste: Inside the 100k-Star Catalog of Agent Design Briefs

Every new application built by an AI coding agent seems to suffer from the same visual condition: a mild, inoffensive, aggressively forgettable sameness. The gradients are soft blue. The cards have rounded corners and generous padding. The typography is Inter, weight 400. The layout works, but it looks like it was assembled by a committee of large language models trained on the same Dribbble top-ten list. As agents like Claude Code and Cursor proliferate, the bottleneck has shifted from implementation to taste. Developers can describe functionality and watch it appear, yet they struggle to make the result look like anything other than a reference implementation of “modern SaaS.”

The VoltAgent/awesome-design-md repository exists to treat this aesthetic monoculture. It has accumulated roughly 100.8k stars and sits at global rank #115 on GitHub, according to star-history tracking. That attention spike is not because the project invents a new framework or optimizes a training pipeline. It is, at its core, a curated list of markdown files. The hook is that these files are design briefs written specifically for machines.

The Sameyness Epidemic

The concept behind the repository originates with Google Stitch, which introduced DESIGN.md as a plain-text companion to the familiar README.md. Where README.md tells an agent how to build the project, DESIGN.md tells it how the project should look and feel. A Threads post summarizing the idea put it succinctly: it is a markdown file that LLMs read to generate consistent UI. The format spread quickly among developers who had grown tired of instructing agents to “make it look like Stripe” and receiving a pastel approximation of a payment form.

The timing is deliberate. The current wave of AI-assisted software development has made it trivial to spin up full-stack applications. What remains difficult is imbuing those applications with a coherent, non-generic visual identity. Design systems have traditionally lived in Figma libraries, JSON token files, and proprietary documentation sites, none of which an AI agent can easily consume during a coding session. The DESIGN.md format collapses that distance. It drops a design system into the project root in the one format LLMs already parse effortlessly: markdown.

A Design System in Plaintext

The repository draws a direct parallel between two files. AGENTS.md is for coding agents; it defines how to build. DESIGN.md is for design agents; it defines how the result should appear. The insight is that markdown is the format LLMs read best, so there is nothing additional to parse or configure. No Figma exports, no JSON schemas, no special tooling. The file is just prose and tables, readable by both human and model.

Each DESIGN.md in the collection follows the Google Stitch specification with extended sections. The structure is methodical: Visual Theme and Atmosphere; Color Palette and Roles with semantic names and hex values; Typography Rules with full hierarchy tables; Component Stylings covering buttons, cards, inputs, and navigation states; Layout Principles including spacing scale and grid philosophy; Depth and Elevation with shadow systems; Do’s and Don’ts that act as guardrails; Responsive Behavior with breakpoints; and finally an Agent Prompt Guide with quick color references and ready-to-use prompts. Every entry also ships with preview.html and preview-dark.html files that render the token catalog as a visual cheat sheet.

This is the technical bet. Rather than forcing an agent to query a vector database or an MCP server for component metadata, the repository gives it a single document that encodes taste, constraint, and decision logic in natural language. It is a low-tech solution to a high-tech problem.

The Catalog of Borrowed Taste

The collection currently contains seventy-two DESIGN.md files extracted from real websites, organized into categories that read like a directory of the modern internet. There are entries for AI platforms such as Anthropic’s Claude, xAI, and Cohere; developer tools including Cursor, Vercel, and Warp; fintech brands like Stripe, Wise, and Revolut; automotive marques from BMW and Ferrari to Renault and Tesla; and media properties including The Verge, WIRED, and Spotify.

The descriptions themselves are agent-ready poetry. VoltAgent’s own entry is summarized as “void-black canvas, emerald accent, terminal-native.” Ferrari’s is “chiaroscuro black-white editorial, Ferrari Red with extreme sparseness.” Lamborghini gets “true black cathedral, gold accent, LamboType custom Neo-Grotesk.” These are not surface-level color pickers. They capture mood, density, and design philosophy in prose that an LLM can reason over.

A Saturday series called “Retro Web · DESIGN.md Nostalgia” pushes the concept further back in time, extracting the visual language of 1996-era sites—literal black page frames, flat color-block ribbon cards, chunky Helvetica-Black over Times Roman—so that an agent can build a period-accurate vintage UI on demand. It is a curatorial choice that treats design history as a recoverable, transferable grammar.

The repository is browsable through a dedicated frontend at getdesign.md, which also offers private extraction requests for sites not yet catalogued. The project is maintained by VoltAgent, a TypeScript framework for building AI agents, and the README does not hide its ecosystem ambitions. It is a community resource, but it is also a marketing surface.

Curation as Infrastructure

To understand where this repository sits in the broader landscape, consider what enterprise teams are building elsewhere. Indeed parsed seventy-seven components from MDX documentation into JSON metadata, ingested them into a Vectra vector database, and benchmarked 1,056 prompts across eight MCP configurations to find the most accurate format. GitHub Primer runs a public MCP server with instruction files and sub-agents that handle accessibility review. Spotify’s Encore design system layers foundations, styles, and behaviors to create smaller context bubbles for AI reasoning.

VoltAgent/awesome-design-md is the lightweight, slightly irreverent counterpoint to all that infrastructure. It is curatorial glue. It reverse-engineers publicly visible CSS values, organizes them into prose, and packages them as a drop-in context window. The README explicitly disclaims ownership of any site’s visual identity and notes that the files are provided “as is” without warranty. These are starting points, not certified design systems. They are fan fiction written in the language of tokens and hex codes.

That honesty is refreshing, but the limits are real. An extracted DESIGN.md captures what a website looked like at the moment of scraping. It cannot enforce runtime consistency the way a live design token pipeline can. It does not version with the brand. And because it is prose-heavy, its precision depends entirely on the quality of the extraction and the reasoning ability of the agent reading it.

The Boring Part Is the Point

The agentic design systems movement argues that AI is a new user of design infrastructure, one that needs structured metadata rather than documentation prose. Advocates at the AI Design Systems Conference 2026 emphasized machine-readable context: JSON schemas per component, deterministic MCP queries, and vectorized knowledge graphs. The underlying fear is that without encoded constraints, an LLM collapses toward the average of the internet.

This repository ignores most of that complexity. It bets that the best way to give an agent context is still plaintext in the format the model reads natively. The boring choice of markdown is exactly why it works. There are no parsers to break, no build steps to fail, no API keys to rotate. A developer copies a file into the project root and the agent’s context window absorbs a complete visual brief. It is a pragmatic implementation of the philosophy that context matters more than probability.

In that sense, the project is less a design system and more a constraint engine. It gives the agent something it cannot retrieve from training data: a specific, opinionated set of encoded decisions about color, type, and spacing. The agent still hallucinates, still drifts, still invents components. But it does so within a narrower aesthetic corridor.

Seeds, Not Trees

The Into Design Systems guide on agentic infrastructure advises teams to “plant seeds, not trees”—to start with naming conventions, token structure, and component descriptions before chasing full autonomy. VoltAgent/awesome-design-md is a seed library. It provides fertile starting soil, not a finished forest.

The open question is whether this format becomes a durable standard or a transitional hack. If major brands begin publishing official DESIGN.md files alongside their Figma libraries and Storybooks, the curatorial work in this repository becomes obsolete in the best possible way. If the agent ecosystem instead moves toward richer, structured formats—JSON component schemas, live MCP servers, or direct Figma API ingestion—then plaintext briefs may look quaint.

For now, the 100.8k stars suggest that developers are eager for a simple, copy-pasteable way to make their agents design like someone, rather than everyone. The repository treats taste as a transferable commodity. It packages the visual languages of the best-designed sites on the web into documents that fit inside a context window. In an era where anyone can prompt their way to a working application, the scarce resource is no longer code. It is the ability to stop the result from looking like it was built by a bot. This collection of markdown files offers one of the simplest possible ways to do that.