Claude organizes with files: the method that avoids chaos

A guide published in 2025 details a method to build a design system with Claude without Figma. The key idea: rely less on a magic prompt and more on living files that frame the model and limit drift.

A guide published in 2025 shows how Claude can be used to build a design system, a component library, and page mockups without using Figma. The principle relies less on a “super prompt” and more on persistent documents that retain project decisions. In this context, the challenge is simple: maintain consistency across sessions without letting the tool go off track. Yes, even models need a somewhat firm specification.

The starting point: frame Claude, don’t vaguely inspire it

The guide emphasizes a very structured initial phase. The user begins by defining the project, the audience, the brand personality, colors, fonts, and design constraints, then requests a foundational HTML/CSS file with tokens, typographic scale, buttons, responsive grid, and section examples.

The idea is not to let Claude “imagine” the style, but to give it composition and rendering rules. According to the author, Claude excels when operating within a precise framework; it becomes weaker as soon as it is asked to create an aesthetic from almost nothing.

Why three documents are better than one long prompt

The proposed method then relies on three files: PROJECT-RULES.md, CURRENT-STATE.md, and ROADMAP.md. The first centralizes all design decisions, corrections, and approved rules. The second tracks the state of the work, what is done and what remains to be confirmed. The third describes the next steps of the work plan.

At each new conversation, Claude must reread these three documents. Without this ritual, the guide observes drift after two or three iterations: styles slipping, components deforming, or decisions forgotten. The system resembles external memory but with the discipline of a project manager who doesn’t let go.

Pages built step by step, not in one block

For each new page, the method first recommends rereading the reference files, then producing a content plan before any full generation. This point is considered crucial: Claude tends to build a whole page too quickly, which then creates more corrections than time savings.

The guide also distinguishes the uses of the chat interface and Claude Code. The former is better suited for creative direction, HTML/CSS prototypes, writing, and visual decisions. The latter is more adapted to integration in a framework, multi-file operations, repetitive migrations, and terminal tasks.

Integrations that really make a difference

Several connectors are considered useful. The Filesystem MCP allows Claude to read and write directly in project files. The Chrome MCP is used to check what actually displays in the browser, especially when a style stubbornly refuses to cooperate. Figma MCP remains optional if the team needs a back-and-forth with mockups.

The guide also mentions Google Analytics MCP to prioritize migration starting from the most strategic pages. Again, the interest is less spectacular than practical: the model works with data, not assumptions.

Models to choose according to the task

The author recommends Opus for design work, as it better maintains visual coherence and aesthetic decisions over time. Sonnet is reserved for mechanical tasks such as content migration, repetitive generation, or data formatting.

In case of a bad start with the wrong model, the advice is straightforward: it’s better to start over than to persist in fixing a shaky result. A useful reminder, especially when time saved at the start turns into visual debt later.

What this method changes for web teams

The guide does not present Claude as a substitute for the designer or project manager role. It rather describes it as an execution accelerator, useful if the team already knows what it wants. The value then comes from continuity: same rules, same vocabulary, same constraints, same source of truth.

In practice, this approach may interest product teams, web studios, and freelancers who build sites at high speed while wanting to avoid dispersion. It also highlights a very concrete point: AI produces better when it inherits a clean system than when it is asked to guess the right direction by chance.

Key points

  • 2025: a guide details a design system method with Claude.
  • Three files structure the project’s memory.
  • Claude chat and Claude Code serve different tasks.
  • Opus is recommended for design, Sonnet for mechanical tasks.
  • MCPs facilitate files, browser, and analytics data.
  • Decisions must be updated after each session.

To read