Ask Claude Code or Cursor to add a pricing section to a Next.js site and the build passes. The section renders. It can still be wrong in ways nothing reports.
The colors come from the default palette instead of your tokens. A second component draws a layout you already have. In a CMS-backed site the section can be missing from the editor's insert menu, so nobody on the content team can ever use it.
A prompt cannot carry everything that prevents this. A new developer gets it in week one: the design rules, the brand rules, the shape of the content model, and a reviewer for their first pull requests. An agent needs the same.
Here is how that looks in a repo you can clone.
Three files the agent reads first
Every skill in the repo starts from the same three files in the project root:
-
DESIGN.mdsays how it looks: tokens, type scale, section anatomy, do's and don'ts. -
BRAND.mdsays what it says: audience, voice, messaging. -
AGENTS.mdsays how to build: architecture, commands, conventions and step-by-step checklists.
In the Meridian demo theme, DESIGN.md has machine-readable front matter with the real token values: colors in OKLCH, the type scale, the fonts. The agent does not guess "a dark gray". It reads dark-card and uses it.
A skill that first asks whether anything needs building
The create-section skill does not start with code. Phase 0 is called "Do we already have it?".
The agent lists the existing sections from SECTION_META, compares the request with the two or three closest types by slots, fields and mechanics, and shows you a short table. Then it decides with you:
- Reuse as-is. An existing type with a layout option or color theme already draws the request. Nothing gets built.
- Extend. An existing type covers most of it. A new optional field or layout variant is added, and existing documents must render exactly as before.
- New type. Nothing fits, so all phases run.
The skill also warns about its own failure mode. Several steps fail silently: a section can compile and typecheck while being invisible in the insert menu, blank in the gallery, or stripped from the document by a stale Studio. So no phase is skipped because the page looks right.
Reviewers that say no
The skill hands off to read-only agents in .agents/agents/. They never edit files. Each returns a verdict (PASS or ISSUES FOUND) and findings in one fixed format: severity, file and line, the rule, the fix.
-
styling-fidelity-reviewerchecks changed code againstDESIGN.md. It flags arbitrary Tailwind values, raw palette classes instead of semantic tokens, and a trap a build never shows: a customtext-NNsize next to atext-{color}insidecn()is silently dropped by tailwind-merge. -
sanity-schema-reviewerchecks the schema. -
groq-parity-reviewerchecks that the GROQ query matches the schema. -
content-model-architectdecides whether something should be an inline object, a document-backed listing or a shared section. -
link-reviewerchecks links and CTAs when a section has them.
The loop is one a team can live with. Marketing asks an agent for a section. The skill reuses what exists, builds only what is missing, and the reviewers check the result against the design and the schema. A developer reviews the final version before anything goes live.
Try it
You need Node.js 22.12+, pnpm and a free Sanity account.
npm create sanity@latest -- --template Pagepro/nexity/base
cd <your project folder>
pnpm dev
pnpm agent:doctor
The site runs on localhost:3000 and the Studio on localhost:3000/admin. pnpm agent:doctor checks that the agent layer is linked for Claude Code and Codex. A live demo with no setup: https://meridian.nexity.build
Pagepro
/
nexity
Nexity is an open source system for starting and growing AI-native websites on Next.js + Sanity.
Nexity
Nexity is an open source system for starting and growing AI-native websites on Next.js + Sanity.
It is for teams where every website change lands on a developer. Nexity gives coding agents (Claude Code, Codex, Cursor) the design system, the blocks and the Sanity schema, so marketing can build new sections with AI and a developer reviews the result before it goes live.
Live demo: https://meridian.nexity.build
Quick start
Requirements: Node.js 22.12 or newer, pnpm 12.15 (corepack enable picks it up), a Sanity account (the free plan works).
npm create sanity@latest -- --template Pagepro/nexity/base
cd <your project folder>
pnpm dev
The site runs on http://localhost:3000 and the Studio on http://localhost:3000/admin.
The project comes with the agent layer: skills and reviewers in .agents/, already linked for Claude Code and Codex. Check the setup with pnpm agent:doctor. To link Cursor, take a newer version of the…
If it saves you a few hours, a star helps other developers find it.
What does your agent get wrong most often in a codebase with a design system: tokens, structure or the content model?

Top comments (0)