Convert design handoff bundles from claude.ai/design (or other design-to-code tools) into a production-grade enterprise Next.js codebase using App Router, Tailwind, TanStack Query, TanStack Table, Zustand, Zod, React Hook Form, and TypeScript strict. The skill supports a dual-mode adapter pattern — a demo mode where every API call reads from local JSON fixtures, and a production mode that hits a real backend (typically Nest.js), switched by NEXT_PUBLIC_API_MODE. Use this skill any time the user mentions a design.html, a handoff bundle, design tokens, converting a prototype to Next.js/React, or "design to code". Use it even with just one HTML file, screenshots, or a verbal description — the skill walks through inventory, token extraction, module planning, scaffolding, dual-mode adapter setup, and module-by-module code generation. Especially trigger when the user mentions enterprise structure, demo mode, JSON mocks, TanStack patterns, or Tailwind token extraction.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Convert design handoff bundles from claude.ai/design (or other design-to-code tools) into a production-grade enterprise Next.js codebase using App Router, Tailwind, TanStack Query, TanStack Table, Zustand, Zod, React Hook Form, and TypeScript strict. The skill supports a dual-mode adapter pattern — a demo mode where every API call reads from local JSON fixtures, and a production mode that hits a real backend (typically Nest.js), switched by NEXT_PUBLIC_API_MODE. Use this skill any time the user mentions a design.html, a handoff bundle, design tokens, converting a prototype to Next.js/React, or "design to code". Use it even with just one HTML file, screenshots, or a verbal description — the skill walks through inventory, token extraction, module planning, scaffolding, dual-mode adapter setup, and module-by-module code generation. Especially trigger when the user mentions enterprise structure, demo mode, JSON mocks, TanStack patterns, or Tailwind token extraction.
Claude Design → Next.js Enterprise Conversion
A pipeline for turning Claude Design HTML output into a production-grade Next.js application. The skill operates in five phases. Walk them in order; skipping phases produces the throwaway code most "design-to-code" pipelines generate.
When to use this skill
Use whenever the user has Claude Design output (HTML, screenshots, a handoff .zip, or a Claude Code handoff folder) and wants a Next.js codebase that is actually maintainable. The trigger is the combination of (a) Claude Design as the input source and (b) Next.js/React as the target. If either is absent, this skill is the wrong tool.
Do NOT use this skill for:
Converting Figma → React (different input shape; no design system extraction step from Claude Design's tokens)
Building a Next.js app from a written spec/PRD with no design artifacts (use the design tooling first, or scaffold from the spec directly)
One-off HTML snippet → single React component (overkill; just do it)
How to invoke this skill
This is a recipe skill — it provides the patterns and references that builder agents follow. Three invocation paths:
Pattern 1 — invoked by the orchestrator (most common)
When prd-design-build-orchestrator runs, its frontend-module-builder subagent reads this skill's references directly. No separate invocation needed; the orchestrator's Phase A.1 install puts this skill at ~/.claude/skills/design-to-nextjs/, and the builder reads SKILL.md + relevant references per the module it's building.
Pattern 2 — invoked by the migration skill
When prototype-to-saas runs Phase 4.0 (bootstrap), the monorepo-bootstrapper agent reads this skill's references/scaffolding.md to set up shadcn correctly in apps/web/. Phase 5's frontend-rewirer reads references/tanstack-patterns.md and references/dual-mode-adapter.md for rewiring patterns.
Pattern 3 — standalone (frontend-only build)
For frontend-only builds without the orchestrator (e.g., converting a Claude Design output into a Next.js codebase you'll wire to an existing backend), start a Claude Code session:
I have a Claude Design handoff at design/index.html and screenshots in design/screenshots/.
Build a production Next.js codebase from it.
The main session reads this skill's SKILL.md and walks the five phases (inventory, tokens, planning, scaffolding, module generation). No subagents required for this path; the main session does the work.
Target stack
The output is always:
Next.js 14+ with App Router, TypeScript strict mode, src/ directory
Tailwind CSS 3.4+ with tokens extracted from the design output
TanStack Query v5 (@tanstack/react-query) for all server state
TanStack Table v8 (@tanstack/react-table) for every data table in the design
Zustand 4+ for client/UI state only (never server state)
Zod for runtime validation of API responses + form schemas
React Hook Form for all forms (paired with Zod resolvers)
shadcn/ui for primitive components (Button, Input, Dialog, etc.) unless the user explicitly opts out
lucide-react for icons
If the user requests a different state library (Jotai, Redux Toolkit), tooltip library, or icon set, swap the relevant module — everything else stays.
Architectural commitments (non-negotiable)
These are baked into every generated app. They are not configurable — if the user wants a different shape, push back and explain why.
1. Monorepo with shared contracts
The Next.js app lives at apps/web/ in a pnpm workspace. Zod schemas and TypeScript types come from packages/contracts (workspace dep @<scope>/contracts), which is shared with the Nest.js backend (apps/api/). The frontend never defines its own response schemas — it imports them.
<workspace>/
├── apps/
│ ├── web/ ← This skill's output
│ └── api/ ← nestjs-enterprise-backend skill's output
├── packages/
│ └── contracts/ ← Zod schemas + view types — SINGLE SOURCE OF TRUTH
If the user is starting fresh, scaffold the monorepo first (using the procedure in the nestjs-enterprise-backend skill's monorepo-setup.md), then run this skill inside it. If the user already has a monorepo, scaffold the Next.js app at apps/web/.
2. View-shape contract — render without ?. or ??
The backend returns view-ready data. Frontend code never uses ?. or ?? to defend against missing fields. This is non-negotiable.
What that means in practice:
Counts are always numbers, defaulted to 0 on the backend. company.counts.contacts is a number, not number | undefined.
Labels are pre-built server-side. A growth signal returns as { kind: 'growing', label: '+12% YoY' } — render signal.label, don't compute.
Variations are discriminated unions.last_activity is { kind: 'None' } | { kind: 'Email Sent', label: string } | .... Pattern-match on kind; don't check for null.
Dates are ISO 8601 strings. Convert to Date at the render site if needed; the type is never Date | undefined.
Genuinely nullable fields are explicit.domain: z.string().nullable() is fine; domain: z.string().optional() is a smell.
The Zod schemas in @<scope>/contracts enforce this — they don't use .optional() on data fields, only on filter inputs. If the user adds an optional field to a view schema, push back.
Frontend reviewers should grep for ?. and ?? in JSX and ask: "Is this defending against a contract gap, or genuinely null-safe?" If the former, fix the contract.
3. Dual-mode adapter (demo + production)
Every generated app supports two modes, switched by NEXT_PUBLIC_API_MODE:
demo — every API call reads from local JSON fixtures under src/mocks/, served by a Next.js Route Handler at /app/api/mock/[...path]/route.ts. No backend required.
production — every API call hits the Nest.js backend at NEXT_PUBLIC_API_BASE_URL.
The production code path is identical between the two modes — same hooks, same Zod validation (against @<scope>/contracts schemas), same error handling. Only the base URL differs.
JSON fixtures live under apps/web/src/mocks/<module>/ and validate against the same @<scope>/contracts schemas the backend's presenters produce. A CI script (pnpm validate:fixtures) runs every fixture through companyViewSchema.parse(fixture) — drift fails the build. See references/dual-mode-adapter.md.
4. Title Case for every enum value — render directly, no conversion
Every enum string in @<scope>/contracts (statuses, stages, roles, industries, discriminator kind fields) is Title Case. Components render company.status directly — no .toUpperCase(), no STATUS_LABELS lookup, no humanize() helper.
If you find yourself reaching for a capitalize/humanize/labelize helper on contract data, the contract is wrong — fix the backend value, not the frontend.
Numeric ranges like "1-10" and "51-200" (size_bucket) are not word-based and render naturally — append units in the view if needed ({company.size.bucket} employees → "51-200 employees").
Fixture files in src/mocks/ use the same Title Case values byte-for-byte. The validate:fixtures script catches drift.
5. shadcn/ui is the ONLY visual primitive library
Every primitive in the app — buttons, inputs, selects, dialogs, tables, cards, badges, dropdowns, tooltips, popovers, sheets, drawers, sidebars, command palettes, toasts — comes from shadcn/ui via @/components/ui/*. The scaffolding step runs pnpm dlx shadcn@latest add ... for every primitive plus the sidebar block up front, so when feature modules are built they import directly without re-installing.
Hand-rolled primitives are also banned. If you find yourself writing a 30-line <MyButton> that wraps <button> with Tailwind classes, stop — use shadcn's <Button>.
Every sidebar uses shadcn's sidebar block (<Sidebar>, <SidebarProvider>, <SidebarTrigger>, <SidebarMenu>, <SidebarMenuItem>, <SidebarHeader>, <SidebarContent>, <SidebarFooter> from @/components/ui/sidebar). No custom <aside className="w-64 ..."> constructions in layout.tsx. If the design shows a sidebar variant the block doesn't natively support (e.g. floating, mini-mode), use the block's built-in variants (collapsible="icon", variant="floating", etc.) rather than re-implementing.
Raw HTML primitives in component code are forbidden where a shadcn equivalent exists: no <button>, <input>, <select>, <textarea>, <dialog>. Layout elements (<div>, <section>, <nav>, <ul>, <li>) are fine — shadcn doesn't ship those.
The exception is the rare case where shadcn genuinely doesn't have something. shadcn covers ~50 primitives — almost everything. If something is missing, surface as a question rather than reaching for an alternative library.
This commitment is enforced by the ui-auditor agent at every wave gate (in orchestrator mode) and as a standalone grep when this skill runs alone — see the validation checklist.
The five-phase pipeline
Phase 1: Inventory → Read every HTML file, screenshot, note. Catalog everything.
Phase 2: Token Extraction → Pull design tokens into a Tailwind config + tokens file.
Phase 3: Module Planning → Split the design into bounded feature modules. Get user buy-in.
Phase 4: Scaffolding → create-next-app, install deps, set up providers + base files.
Phase 5: Module Generation → For each module: types → schemas → api → hooks → store → components → page.
Walk them in order. Phase 3 has a mandatory user-confirmation gate (don't generate code against a module plan the user hasn't seen). Phases 4 and 5 can be partially parallelized but should be presented to the user sequentially.
Phase 1 — Inventory the input
Goal: Build a complete catalog of what's in the Claude Design output before writing any code.
Step 1.1 — Locate and inspect. Use view on the folder if it's a directory, then read every HTML file. If there's a design-notes.md or README.md, read those first — they contain the PM's intent. Do not skim these. The design notes often contain critical information that the HTML doesn't (e.g., "this list is server-paginated", "the table should support multi-select").
Step 1.2 — Catalog the screens. A "screen" maps roughly 1:1 to a Next.js route. For each HTML file (or each major <section> block in a single-file design), record:
Screen name (best guess; confirm with user later)
Auth requirement (login screens are public; dashboards are gated)
Whether it's a list, detail, form, or composite view
Step 1.3 — Catalog repeated patterns. Walk each HTML file and identify components that appear more than twice. These are the candidates for components/ui/ or module-scoped components. Common examples in CRM/SaaS designs: card, stat tile, table row, badge, avatar, empty state, toolbar, filter chip, side drawer.
Step 1.4 — Catalog tables. Every <table> or grid-like structure becomes a TanStack Table. For each, record:
Columns (label, what data it shows, sortable/filterable hints from the design)
Whether the design implies server-side or client-side pagination
Selection model (single, multi, none)
Row actions (the ... menu, inline buttons)
Step 1.5 — Catalog forms. Every form becomes a React Hook Form + Zod schema. For each:
Fields (name, type, required, validation hints from the design)
Submit behavior (suggests a mutation)
Multi-step? (suggests a wizard pattern + a Zustand store for cross-step state)
Step 1.6 — Catalog interactive widgets. Drawer/modal/dropdown/tabs/accordion — note them, but don't over-engineer; most will become shadcn/ui primitives.
Step 1.7 — Read design-notes.md for intent that isn't in the HTML. Empty states, loading states, error states, what happens on hover, animation hints, accessibility requirements.
Output of Phase 1: an inventory document. Write it out as INVENTORY.md in the working directory so the user can review it. Keep it terse — bullet lists, no prose.
Phase 2 — Token extraction
Goal: Extract design tokens (colors, typography, spacing, radii, shadows, motion) from the Claude Design HTML into a Tailwind config + a tokens.ts file. This is the single highest-leverage step. Skipping it leads to hex codes scattered across components and a UI that drifts within a week.
See references/token-extraction.md for the full extraction procedure and the canonical tailwind.config.ts template.
Quick summary of what to extract:
Colors — collect every unique hex/rgb/hsl value from inline styles and Tailwind classes (bg-[#1a2b3c], text-[hsl(220_30%_15%)]). Cluster by purpose (brand, surface, text, border, status). Name them semantically (brand.500, surface.muted, text.primary), not by hex.
Typography — font families (look for font-family in inline styles or font-['...'] in classes), font sizes used, weights used, line-heights. Build a type scale even if the design uses arbitrary sizes.
Spacing — Tailwind has a default scale (0, 0.5, 1, 1.5, 2, ...); extend only if the design uses arbitrary values consistently.
Radii — round any inline border-radius values to the nearest Tailwind step or add custom ones.
Shadows — extract every shadow into a named token.
Breakpoints — usually leave as default unless the design uses custom ones.
Motion — durations and easings used in transition and @keyframes.
Output of Phase 2:
tailwind.config.ts (final form, ready to drop in)
src/styles/tokens.ts (raw token values for use in JS/TS)
src/styles/globals.css (CSS custom properties for dark mode / theming if applicable)
Goal: Decide how the design splits into feature modules before any code is generated. Get explicit user sign-off before proceeding.
Why this gate matters: Module boundaries are the single most expensive thing to change later. A feature scoped wrong now becomes a multi-week refactor in three months. Always show the proposed split to the user and wait for confirmation.
See references/enterprise-structure.md for the canonical folder layout and module conventions.
Step 3.1 — Group screens into modules. A module is a bounded business capability. For an <APP_NAME>-like product, typical modules:
auth — login, signup, password reset, MFA
workspace — workspace settings, members, billing
companies — list, detail, import, custom fields
contacts — list, detail, conversation history
campaigns — list, builder, preview, send
inbox — unified inbox, thread view, reply
pipeline — kanban board, lead detail, deal conversion
analytics — dashboards, reports
The right module count for most products is 5–10. Fewer than 5 = monolith risk. More than 12 = over-fragmented.
Step 3.2 — Identify shared primitives. What lives in components/ui/ (shadcn primitives) vs components/shared/ (cross-module composites like <DataTable>, <PageHeader>, <EmptyState>).
Step 3.3 — Identify cross-module dependencies. A Lead references a Contact references a Company. Decide whether modules expose typed public APIs (modules/companies/index.ts re-exports types + hooks) or whether everything imports from deep paths. Default to typed public APIs — it forces clean boundaries.
Wait for confirmation before moving to Phase 4. If the user pushes back, iterate. Do not skip this gate even if the user seems impatient — a wrong module split costs more time than the confirmation step does.
Phase 4 — Scaffolding
Goal: Get a runnable Next.js app with all dependencies, providers, and base files in place. The app should pnpm dev cleanly with an empty home page before any feature code is added.
See references/scaffolding.md for the full setup procedure (commands, config files, providers).
Set up the enterprise folder layout (empty directories with .gitkeep):
src/
├── app/
├── components/
│ ├── ui/ ← shadcn primitives go here
│ └── shared/ ← cross-module composites
├── modules/ ← one folder per planned module
├── lib/
├── hooks/
├── stores/ ← global Zustand stores only
├── styles/
└── types/
Configure path aliases in tsconfig.json: @/*, @/modules/*, @/components/*, @/lib/*
Set up ESLint + Prettier + (optionally) Husky + lint-staged
Verify the app boots: empty home page should render
Don't generate feature code until scaffolding boots cleanly. It's much harder to debug missing-provider or path-alias issues once 50 files are in flight.
Phase 5 — Module-by-module generation
Goal: For each planned module, generate the code in a deterministic order. Resist the urge to start with components — components without types/hooks behind them encode bad assumptions that are painful to undo.
The canonical order within a module:
1. (schemas come from @<scope>/contracts; no local types or schemas)
2. api/ — fetcher functions (getCompanies, createCompany) using the api-client wrapper,
typed against schemas imported from @<scope>/contracts
3. hooks/ — TanStack Query hooks (useCompanies, useCreateCompany, useUpdateCompany)
4. store/ — Zustand store IF AND ONLY IF the module has UI state that crosses components
(e.g., a multi-step form, a selection state shared between toolbar and table)
5. fixtures — JSON files under src/mocks/<module>/ (list.json, detail.json, create.json, ...)
each fixture validates against the module's schema in @<scope>/contracts
6. components/ — feature components, including DataTable column defs.
NO ?. or ?? — render contract fields directly.
7. (pages) — files under src/app/(route-group)/ that import from modules/
See references/tanstack-patterns.md for query/mutation patterns, references/zustand-patterns.md for store conventions, references/component-patterns.md for translating design HTML to React components, and references/dual-mode-adapter.md for fixture authoring rules.
Tactical rules during generation:
Server Components by default. Mark client components with 'use client' only when you need event handlers, hooks, or browser APIs. Lists are server components; tables that filter/sort/paginate are client.
Never use Zustand for server state. TanStack Query owns the cache. Zustand owns ephemeral UI state (modal open/closed, selected row IDs, active tab).
Every API response goes through Zod. The api-client wrapper takes a Zod schema and returns the parsed type. This catches contract drift the moment it happens.
Every form uses React Hook Form + Zod resolver. No raw useState for forms. No uncontrolled inputs.
Every table uses TanStack Table. No raw <table> with manual state. The <DataTable> shared component takes column defs + data + options.
Imports respect module boundaries. Within a module, deep imports are fine. From outside, import only from the module's index.ts.
Order across modules: start with the module that has the most cross-module dependents (usually companies or auth for a CRM). Build it end-to-end (types → page renders with mock data). Then build the next module that depends on it. This produces working slices instead of half-built layers.
Validation checklist
Before declaring done, walk this list:
pnpm typecheck passes (TypeScript strict, no any)
pnpm lint passes (zero warnings, not just zero errors)
pnpm build succeeds
pnpm validate:fixtures passes (every JSON fixture conforms to its @<scope>/contracts Zod schema)
No ?. or ?? in src/modules/**/components/ — grep and review every hit. If it's defending against a contract gap, fix the contract; if it's genuine null safety on a nullable field, leave it.
No local Zod schemas in apps/web/src/ — all schemas come from @<scope>/contracts. Grep for z.object outside node_modules; non-form schemas are a smell.
No enum casing helpers in components — grep for capitalize, humanize, toLowerCase, toUpperCase, LABELS[, _LABEL inside src/modules/**/components/. Any hit on contract data means a contract value isn't Title Case; fix the contract, not the component.
No competing UI libraries — grep -rEn "from '(@radix-ui/|@headlessui/|@mui/|@material-ui/|@chakra-ui/|@mantine/|antd|@ant-design/|react-bootstrap|flowbite-react|@nextui-org/|tremor|@tremor/|daisyui)" apps/web/src --include="*.tsx" --include="*.ts" | grep -v "apps/web/src/components/ui/" returns ZERO hits
No raw HTML primitives where shadcn equivalents exist — grep -rEn "<button(\\s|>)|<input(\\s|>)|<select(\\s|>)|<textarea(\\s|>)|<dialog(\\s|>)" apps/web/src/modules --include="*.tsx" | grep -v "apps/web/src/components/ui/" returns zero hits (excepting hidden inputs)
Sidebar uses shadcn block — every layout with navigation imports from @/components/ui/sidebar; no custom <aside className="w-64 ..."> in any layout.tsx
No hand-rolled primitives in modules — function MyButton, function MyDialog, etc. that wrap raw HTML
No arbitrary color/radius Tailwind classes — grep -rEn "(bg|text|border)-\\[#[0-9a-fA-F]{3,8}\\]|rounded-\\[" apps/web/src/modules returns zero hits
App boots cleanly in demo mode — every route renders with realistic fixture data
App boots cleanly in production mode when pointed at the Nest.js backend
Every route in the design renders at /[route]
Every table sorts and filters
Every form submits and shows loading + error + success states
Every mutation invalidates the right queries
Dark mode (if in the design) toggles correctly
Responsive: 375px, 768px, 1280px — no horizontal scroll
No hex codes outside tailwind.config.ts and tokens.ts
No console.log left in code
No // TODO left without a tracking issue
If the user provided screenshots, do a final visual diff: open the screenshot and the live page side-by-side, compare spacing, color, typography. Flag any drift.
Reference files
Load these as needed during the phases above. Each is focused; don't read all of them upfront.
Phase 5 (when to use Zustand and the store conventions)
Common pitfalls
These come up every time. Watch for them.
Pitfall 1: Treating Claude Design's HTML as the final markup. It isn't. It's a visual specification. The job is to re-implement against the target stack's components, not to copy-paste HTML into JSX. If you find yourself transcribing <div class="..."> blocks one-for-one, stop and refactor into proper components.
Pitfall 2: Hex codes leaking into components. Every color reference outside tailwind.config.ts and tokens.ts is technical debt. Map them all in Phase 2.
Pitfall 3: Mixing server and client state. Zustand for useCompanies() data is a smell. TanStack Query owns server state. Zustand owns UI state. Confusing these two costs weeks of refactoring later.
Pitfall 4: Skipping Phase 3 confirmation. Generating 50 files against the wrong module split is the most expensive failure mode. Always confirm.
Pitfall 5: Generating components before types/hooks. Reverse the order and components encode wrong assumptions that propagate.
Pitfall 6: Using useEffect for data fetching. Use TanStack Query. Always. useEffect(() => fetch(...)) is a 2020 pattern.
Pitfall 7: Over-using 'use client'. Most pages should be server components. Push 'use client' to the leaves (the interactive widgets), not the page roots.
Pitfall 8: Forgetting to set queryKey consistently. Use a centralized queryKeys object per module (see references/tanstack-patterns.md) so invalidation is predictable.
