How to give Claude Code and Cursor your brand, not just your design system

What a design system leaves out

A design system supplies components and styling. A landing page also needs product claims, voice and approved imagery.

Google Labs’ DESIGN.md describes visual identity to coding agents. It does not supply your product’s claim evidence.

What Claude Code and Cursor read

Claude Code’s memory guide recommends under 200 lines per CLAUDE.md; imports still consume context. Cursor recommends rules under 500 lines and does not inline a referenced file’s contents; a Claude @path import does.

Both treat instructions as guidance, so check results. AGENTS.md also serves other coding tools; import it from CLAUDE.md with @AGENTS.md when both exist. Skills and Cursor skills load when relevant, but a copied brand guide still needs updating.

Keep code conventions in the repo

Keep these with code review:

• Components, accessibility rules and composition patterns.
• Tokens and their meanings.
• Original logos, licensed fonts and icons the build ships.
• Run/preview instructions.

Change the description with the token in the same pull request.

Fetch what changes with the brand

Voice, terms, supported claims and campaign direction also serve work outside this repository. Copying them into every rules file creates another update to make when the brand changes.

Use this distinction: if a change belongs in this repository’s pull request, keep it here; if it belongs to the shared brand, retrieve it for the task.

Write checkable rules and a usable voice guide rather than adjectives.

Connect a maintained brand

Rainbrand maintains shared identity, guidance and assets. Connect Claude Code or Cursor, confirm the account and brand, then request a Brief for the page.

The agent builds with your stack. In the Mercury Codex run, it downloaded the supplied Arcadia fonts and original logo for a payment-confirmation concept.

The first colour check failed the card background. Codex corrected it; logo findings also required judgment, including one conflicting with the Brief and a later spacing finding that appeared wrong. The final image passed the four submitted checks. The saved run retains those findings.

For shared setup, commit the server address, not a credential:

The URL is https://mcp.rainbrand.com/mcp. Each person signs in with their own Rainbrand account; add the brand to the intended team. Claude Code asks for approval before first use; Cursor documents its MCP setup.

After a reviewed, applied brand edit, new Briefs use current guidance. Shipped files still change through a pull request. Scripts and services can use the API.

Tell the agent when to fetch it

Put short retrieval instructions in AGENTS.md:

• Request the selected brand’s Brief before writing copy or choosing images.
• Build with repository components and tokens.
• Use only supported claims; leave a TODO for missing evidence.
• Import original logos instead of redrawing them.
• Preview desktop/mobile and report unresolved findings.

To scope those instructions, use Claude .claude/rules/ with paths, or Cursor .cursor/rules/ with globs. A matching rule may load after a page starts; short always-loaded retrieval instructions avoid that delay.

Request a Rainbrand Brief for [your brand] for the page described in the attached task. Build it with this project’s components and tokens, using the Brief’s guidance for the copy, its selected images and the original logo file. Keep the audience and call to action from the task, and make only claims supported by the supplied product sources. Preview it at desktop and mobile widths, request Reviews of the finished captures with the expected copy, then show me the page and unresolved findings.

Check the page before it merges

Use Rainbrand Review on finished captures with expected copy and constraints. Revise confirmed findings and review updated captures; image checking covers that loop.

Try the web-page workflow on your next real page. Keep source/code checks alongside the brand findings.