Design.md Generator: A skill to codify design taste

There's a pattern repeating in the Claude ecosystem since skills and MCP servers became standard practice: someone identifies a communication problem between humans and models, converts it into a structured file, and proposes it as a de facto standard. This week, a concrete example appeared on Hacker News: AI Design Taste – Design.md Generator, a Chrome extension that analyzes a website and generates a `design.md` file with its design principles encoded.

The proposal is straightforward: instead of describing your aesthetic preferences in each conversation or prompt, you produce a structured document—palette, typography, visual density, spacing rules—that you can include as persistent context in Claude Code, as part of a skill, or simply attach to any conversation. The Hacker News discussion has limited traction so far (1 point, no comments), which says something about how early the project is.

What it does exactly

The extension installs in Chrome, points to any URL you choose as a design reference, and extracts—presumably through DOM analysis and computed styles—a set of design decisions that it outputs to a markdown file. The result is something like a `design.md` that describes:

• Primary typeface family and type scale
• Color palette with its roles (primary, background, text, accent)
• Spacing patterns and border radius
• Overall visual tone (minimalist, dense, expressive, etc.)

That file can then be injected as context into any LLM. In the Claude Code workflow, it would fit well as a design skill loaded at the start of a frontend project, or as part of the context for a specialized subagent focused on UI component generation.

Why it makes conceptual sense

The problem it tackles is real. When you ask Claude to generate a UI component, the model has its own judgment but no access to your preferences or your project's visual conventions. You can describe the style in the prompt, but that's fragile: it changes between sessions, gets forgotten, and often the verbal description doesn't capture precisely what you have in mind.

The idea of externalizing that knowledge into a reusable artifact—a `design.md` that's versionable in your repository and loadable as a skill in Claude Code—aligns with how modern context engineering works. It's not that different from what many people already do with `CLAUDE.md` for project rules: a reference document the model consumes before acting.

Claude Opus 4.7's 1M token window makes adding that extra context essentially free in terms of token cost. The limit is no longer what fits, but what's worth including.

Who it's useful for (and who it isn't)

The clearest use case is the frontend developer working solo or in small teams, without a formal design system, who wants Claude to stay consistent with the visual style of an existing product. In that scenario, generating a `design.md` from your production website and including it in your working context is a reasonable shortcut.

Less useful, or even counterproductive, for teams that already have a documented design system with well-defined tokens (Figma Tokens, Style Dictionary, etc.). There, `design.md` would be an additional abstraction layer with risk of drifting from the actual system.

There's also a legitimate question about analysis quality. Extracting design criteria from the DOM and computed styles is imprecise: it captures what's rendered, not the designer's intentions or contextual exceptions. A website with technical debt can produce a `design.md` that encodes exactly the mistakes you'd want to avoid.

Project status

As of publication on Hacker News on May 9, 2026, the extension is available in the Chrome Web Store but there are no signs of public documentation, an open repository, or native integrations with Claude Code or MCP servers. It's an early experiment, not production-ready tooling.

---

The direction is interesting: turning design taste into a portable, machine-readable artifact is an unsolved problem. Whether the current solution—a Chrome extension scraping CSS styles—is the right approach remains to be seen, but the concept deserves more attention than it's getting on HN.