PARLOR SITE FACTORY

From Description
to Deployed Website.

A 7-phase agentic pipeline that uses Gemini 3.1 Pro to generate complete, pixel-perfect, multi-page websites from a URL or business description. No templates. No themes. Pure AI generation.

See the Pipeline ↓

7 PHASES

The Complete Pipeline

Every phase is handled by a specialized AI agent. Each one builds on the last.

Intake

Claude Sonnet

In: URL or business description
Out: BusinessIntelligence JSON

Scrapes the URL, strips HTML, extracts: name, industry, products, tone, booking system, competitor references.

Design System

Gemini 3.1 Pro

In: BusinessIntelligence
Out: Locked DesignSystem JSON

Generates 4-color palette with semantic names, 3-font system (heading + emphasis serif + data mono), border radius, noise texture, animation easing curves. CONTEXT CACHED — 90% token savings on all pages.

Site Architecture

Gemini 3.1 Pro

In: BusinessIntelligence + DesignSystem
Out: SiteMap JSON

Industry-adaptive: restaurants get menu + reservations pages, SaaS gets pricing + demo sandbox, legal gets practice areas + compliance + intake.

Mega-Prompts

Claude Sonnet

In: SiteMap + DesignSystem
Out: One mega-prompt per page

3-phase prompt structure: architecture → UI/UX components → trust/compliance. Section templates from 48-template library. Live data requirements injected per page (menus, availability, blog posts, reservations). Negative constraints always last. Execution directive at end.

Page Generation

Gemini 3.1 Pro

In: Mega-prompts (all pages parallel)
Out: Complete Next.js page components

All pages generated in parallel. Design system cached across calls. Live data hooks (useLiveData) wired per page — real-time menus, availability, and CRM feeds bound at generation time. Generates design-tokens.ts on first page. No hardcoded values.

Module Injection

Claude Haiku

In: Generated pages
Out: Pages with modules wired

Surgically injects four design-system-native modules: Kit (AI chat widget), Scribe (blog engine), CRM (lead capture), Reservations (4-state booking machine: party size → date/time → contact → confirmation — supports Resy, OpenTable, phone, and walk-ins).

QA Agent

Gemini 3.1 Pro (multimodal)

In: Screenshots + source code
Out: CSS patches + quality score

Gemini reviews its own rendered output. Finds layout breaks, contrast failures, generic AI patterns. Target: 90% quality score. Re-screenshots after patches to verify.

Delivery

—

In: QA-verified pages
Out: Live website

One-click deploy to Cloudflare Pages. Page-by-page preview gallery. Custom domain support.

PHASE 1 OUTPUT

The Design System — Global Contract

Generated once, context-cached, injected into every page prompt. No page ever hardcodes a color or font name.

Burnt Sienna

#9B4C4C

Primary CTAs

Gold Leaf

#D4AF37

Accents, hover

Cream Linen

#F7F5F0

Background

Deep Charcoal

#2A2A2A

Text, footer

Warm Taupe

#A8937A

Muted text

// lib/design-tokens.ts — auto-generated, imported by every page

export const colors = {
  primary:    '#9B4C4C',   // Burnt Sienna  — CTAs, highlights
  accent:     '#D4AF37',   // Gold Leaf     — accents, hover states
  background: '#F7F5F0',  // Cream Linen   — page background
  dark:       '#2A2A2A',  // Deep Charcoal — text, footer
};

export const fonts = {
  heading:  'Plus Jakarta Sans',   // structure, hierarchy
  emphasis: 'Cormorant Garamond',  // italic serif — drama
  data:     'JetBrains Mono',      // metrics, timestamps
};

export const easing = {
  spring:   'cubic-bezier(0.34, 1.56, 0.64, 1)',
  reveal:   'cubic-bezier(0.16, 1, 0.3, 1)',
  duration: '0.6s',
};

PHASE 3 OUTPUT

One Mega-Prompt Per Page

Every page gets a Nura-Health-quality prompt assembled from section templates. This goes to Gemini 3.1 Pro and produces a complete Next.js component.

// HOME PAGE MEGA-PROMPT — Cucina Verona (restaurant example)

# HOME — Cucina Verona

## ROLE & OBJECTIVE
You are a World-Class Senior Creative Technologist and Lead Frontend Engineer.
Build the Home page for Cucina Verona, a Northern Italian restaurant in Hayes Valley SF.
Aesthetic identity: "Rustic Modern / Italian Precision" — NOT generic AI design.

## CORE DESIGN SYSTEM (LOCKED — CONTEXT CACHED)
Palette: Burnt Sienna #9B4C4C (primary) · Gold Leaf #D4AF37 (accent)
Fonts: Plus Jakarta Sans (heading) · Cormorant Garamond italic (emphasis)
Easing: cubic-bezier(0.34, 1.56, 0.64, 1) for spring animations
Import all values from lib/design-tokens.ts — never hardcode.

## PHASE 1 — CORE ARCHITECTURE
- Semantic HTML5 · WCAG 2.2 AA · Schema.org LocalBusiness JSON-LD

## PHASE 2 — UI/UX COMPONENTS

### HERO — The Immersive Table
100dvh · Forest/Italian scene background (real Unsplash URL) · Moss-to-black gradient
Bottom-left content · "Welcome to" (Plus Jakarta Sans bold) / "Cucina Verona." (Cormorant italic, massive)
GSAP stagger: eyebrow 0.2s → headline 0.5s → sub-copy 0.9s → CTAs 1.1s

### FEATURES — The Precision Micro-UI Dashboard
3 interactive cards (feel like running software):
  Card 1: Tonight's Specials Shuffler — 3 overlapping cards rotate every 3s via unshift(pop())
  Card 2: Live Availability — pulsing dot + time slot pills + typewriter messages
  Card 3: Reservation Scheduler — automated SVG cursor clicks through week grid

### PHILOSOPHY — The Manifesto
Dark section · parallax texture · split-text GSAP reveal:
  "Modern restaurants serve food." (50% white, small)
  "We serve experiences." (Gold Leaf, 2x larger, bold italic)

### MENU PREVIEW
5 dishes · category tabs · image hover scale · price in accent color

### BOOKING WIDGET
4-state machine: party size → date/time → contact → confirmation
Resy integration · POST /api/cucinaverona/reservations

## LIVE DATA REQUIREMENTS
This page binds live data via useLiveData() hook:
  menu_items  → /api/cucinaverona/menu?category=featured (refreshes every 5min)
  availability → /api/cucinaverona/availability (real-time, refreshes every 60s)
Show skeleton states while loading. Animate in with spring easing on resolve.

## PHASE 3 — MODULES
Kit: floating chat icon (60px) → 380x600px modal, streams from /api/chat
Scribe: 3 recent blog posts as cards (fetch from /api/cucinaverona/blog)
Reservations: 4-state machine (party size → date/time → contact → confirmation)
  POST /api/cucinaverona/reservations · Resy integration · design-system native

## NEGATIVE CONSTRAINTS ← ALWAYS LAST
DO NOT use generic AI design (purple gradients, system fonts, flat cards)
DO NOT use placeholder images — real Unsplash URLs only
DO NOT use low-contrast text — 4.5:1 WCAG minimum
DO NOT implement GSAP without gsap.context() cleanup

## EXECUTION DIRECTIVE
Do not build a page. Build a digital instrument.
Every scroll should feel intentional. Eradicate all generic AI patterns.

PHASE 5 OUTPUT

Design-System-Native Modules

Four production-grade modules injected into every generated site. All inherit the generated design system — no theming required.

Kit

AI Chat Widget

kit.ts

Floating 60px icon → 380x600px modal. Streams from /api/chat. Design-system themed. Context-aware of the business and current page.

Scribe

Blog Engine

scribe.ts

Fetches from /api/[slug]/blog. 3-card preview on site pages, full CMS on /blog. AI-generated posts indexed for SEO and AEO.

CRM

Lead Capture

crm.ts

Inline lead capture forms wired to Supabase. Auto-tags leads by page source. Feeds into tenant CRM dashboard.

Reservations

4-State Booking Machine

reservations.ts

1. Party size → 2. Date & time slot → 3. Contact info → 4. Confirmation. Supports Resy, OpenTable, phone, and walk-in flows. Posts to /api/[slug]/reservations.

The Key Insight

One locked design system. Live data wired at generation. Four production modules injected.

The design system is generated once and context-cached — every page loads it at ~10% of the normal token cost. Live data requirements (menus, availability, blog, reservations) are injected directly into each page's mega-prompt so the AI generates real data bindings — not static placeholders. All pages run in parallel. Modules inherit the design system automatically. The result: a cohesive, pixel-perfect, living multi-page site where nothing is hardcoded and everything scales.

RESEARCH FOUNDATION

Built on Proven Best Practices

Best Practice	Why It Matters	Applied How
thinking_level=HIGH	Allocates max tokens for internal reasoning before code generation	Set on every Gemini 3.1 Pro call
temperature=1.0	Lowering temp causes instability and degraded logic in Gemini 3.1 Pro	Never changed — locked at 1.0
Context Caching	90% token reduction for cached design system	DesignSystem JSON cached, reused across all pages
3-Phase Mega-Prompt	Separates architecture, UI/UX, compliance concerns	Phase 1 (HTML), Phase 2 (components), Phase 3 (modules)
Negative Constraints Last	Model retains DO NOT rules through lengthy generation	Always final block in every prompt
Multimodal QA	Model debugs its own visual output via screenshot review	Screenshot → Gemini → CSS patches loop
WCAG 2.2 AA	Mandatory accessibility for all 2026 web builds	aria-labels, 4.5:1 contrast, keyboard nav in every prompt
Schema.org JSON-LD	AEO/GEO: AI crawlers need structured data, not just SEO	Every page includes LocalBusiness or industry-specific schema

EXAMPLE

Italian Restaurant → Live Website

Cucina Verona, Hayes Valley SF — from description to deployed in ~5 minutes

INPUT

name: "Cucina Verona"
industry: restaurant
personality: intimate, rustic-modern
booking: resy
modules: kit, scribe, reservations
competitors: Quince, SPQR

OUTPUT

✓ Design System JSON (cached)
✓ SiteMap: 4 pages, 24 sections
✓ 4 mega-prompts + live data hooks
✓ 4 Next.js page components
✓ design-tokens.ts generated
✓ Live data bound (menu, availability)
✓ Kit + Scribe + CRM + Reservations
✓ QA score: 89% (all pass)
✓ Live on Cloudflare Pages

PARLOR SITE FACTORY

From Description
to Deployed Website.

A 7-phase agentic pipeline that uses Gemini 3.1 Pro to generate complete, pixel-perfect, multi-page websites from a URL or business description. No templates. No themes. Pure AI generation.

See the Pipeline ↓

7 PHASES

The Complete Pipeline

Every phase is handled by a specialized AI agent. Each one builds on the last.

Intake

Claude Sonnet

In: URL or business description
Out: BusinessIntelligence JSON

Scrapes the URL, strips HTML, extracts: name, industry, products, tone, booking system, competitor references.

Design System

Gemini 3.1 Pro

In: BusinessIntelligence
Out: Locked DesignSystem JSON

Site Architecture

Gemini 3.1 Pro

In: BusinessIntelligence + DesignSystem
Out: SiteMap JSON

Industry-adaptive: restaurants get menu + reservations pages, SaaS gets pricing + demo sandbox, legal gets practice areas + compliance + intake.

Mega-Prompts

Claude Sonnet

In: SiteMap + DesignSystem
Out: One mega-prompt per page

Page Generation

Gemini 3.1 Pro

In: Mega-prompts (all pages parallel)
Out: Complete Next.js page components

Module Injection

Claude Haiku

In: Generated pages
Out: Pages with modules wired

QA Agent

Gemini 3.1 Pro (multimodal)

In: Screenshots + source code
Out: CSS patches + quality score

Gemini reviews its own rendered output. Finds layout breaks, contrast failures, generic AI patterns. Target: 90% quality score. Re-screenshots after patches to verify.

Delivery

—

In: QA-verified pages
Out: Live website

One-click deploy to Cloudflare Pages. Page-by-page preview gallery. Custom domain support.

PHASE 1 OUTPUT

The Design System — Global Contract

Generated once, context-cached, injected into every page prompt. No page ever hardcodes a color or font name.

Burnt Sienna

#9B4C4C

Primary CTAs

Gold Leaf

#D4AF37

Accents, hover

Cream Linen

#F7F5F0

Background

Deep Charcoal

#2A2A2A

Text, footer

Warm Taupe

#A8937A

Muted text

// lib/design-tokens.ts — auto-generated, imported by every page

export const colors = {
  primary:    '#9B4C4C',   // Burnt Sienna  — CTAs, highlights
  accent:     '#D4AF37',   // Gold Leaf     — accents, hover states
  background: '#F7F5F0',  // Cream Linen   — page background
  dark:       '#2A2A2A',  // Deep Charcoal — text, footer
};

export const fonts = {
  heading:  'Plus Jakarta Sans',   // structure, hierarchy
  emphasis: 'Cormorant Garamond',  // italic serif — drama
  data:     'JetBrains Mono',      // metrics, timestamps
};

export const easing = {
  spring:   'cubic-bezier(0.34, 1.56, 0.64, 1)',
  reveal:   'cubic-bezier(0.16, 1, 0.3, 1)',
  duration: '0.6s',
};

PHASE 3 OUTPUT

One Mega-Prompt Per Page

Every page gets a Nura-Health-quality prompt assembled from section templates. This goes to Gemini 3.1 Pro and produces a complete Next.js component.

// HOME PAGE MEGA-PROMPT — Cucina Verona (restaurant example)

# HOME — Cucina Verona

## ROLE & OBJECTIVE
You are a World-Class Senior Creative Technologist and Lead Frontend Engineer.
Build the Home page for Cucina Verona, a Northern Italian restaurant in Hayes Valley SF.
Aesthetic identity: "Rustic Modern / Italian Precision" — NOT generic AI design.

## CORE DESIGN SYSTEM (LOCKED — CONTEXT CACHED)
Palette: Burnt Sienna #9B4C4C (primary) · Gold Leaf #D4AF37 (accent)
Fonts: Plus Jakarta Sans (heading) · Cormorant Garamond italic (emphasis)
Easing: cubic-bezier(0.34, 1.56, 0.64, 1) for spring animations
Import all values from lib/design-tokens.ts — never hardcode.

## PHASE 1 — CORE ARCHITECTURE
- Semantic HTML5 · WCAG 2.2 AA · Schema.org LocalBusiness JSON-LD

## PHASE 2 — UI/UX COMPONENTS

### HERO — The Immersive Table
100dvh · Forest/Italian scene background (real Unsplash URL) · Moss-to-black gradient
Bottom-left content · "Welcome to" (Plus Jakarta Sans bold) / "Cucina Verona." (Cormorant italic, massive)
GSAP stagger: eyebrow 0.2s → headline 0.5s → sub-copy 0.9s → CTAs 1.1s

### FEATURES — The Precision Micro-UI Dashboard
3 interactive cards (feel like running software):
  Card 1: Tonight's Specials Shuffler — 3 overlapping cards rotate every 3s via unshift(pop())
  Card 2: Live Availability — pulsing dot + time slot pills + typewriter messages
  Card 3: Reservation Scheduler — automated SVG cursor clicks through week grid

### PHILOSOPHY — The Manifesto
Dark section · parallax texture · split-text GSAP reveal:
  "Modern restaurants serve food." (50% white, small)
  "We serve experiences." (Gold Leaf, 2x larger, bold italic)

### MENU PREVIEW
5 dishes · category tabs · image hover scale · price in accent color

### BOOKING WIDGET
4-state machine: party size → date/time → contact → confirmation
Resy integration · POST /api/cucinaverona/reservations

## LIVE DATA REQUIREMENTS
This page binds live data via useLiveData() hook:
  menu_items  → /api/cucinaverona/menu?category=featured (refreshes every 5min)
  availability → /api/cucinaverona/availability (real-time, refreshes every 60s)
Show skeleton states while loading. Animate in with spring easing on resolve.

## PHASE 3 — MODULES
Kit: floating chat icon (60px) → 380x600px modal, streams from /api/chat
Scribe: 3 recent blog posts as cards (fetch from /api/cucinaverona/blog)
Reservations: 4-state machine (party size → date/time → contact → confirmation)
  POST /api/cucinaverona/reservations · Resy integration · design-system native

## NEGATIVE CONSTRAINTS ← ALWAYS LAST
DO NOT use generic AI design (purple gradients, system fonts, flat cards)
DO NOT use placeholder images — real Unsplash URLs only
DO NOT use low-contrast text — 4.5:1 WCAG minimum
DO NOT implement GSAP without gsap.context() cleanup

## EXECUTION DIRECTIVE
Do not build a page. Build a digital instrument.
Every scroll should feel intentional. Eradicate all generic AI patterns.

PHASE 5 OUTPUT

Design-System-Native Modules

Four production-grade modules injected into every generated site. All inherit the generated design system — no theming required.

Kit

AI Chat Widget

kit.ts

Floating 60px icon → 380x600px modal. Streams from /api/chat. Design-system themed. Context-aware of the business and current page.

Scribe

Blog Engine

scribe.ts

Fetches from /api/[slug]/blog. 3-card preview on site pages, full CMS on /blog. AI-generated posts indexed for SEO and AEO.

CRM

Lead Capture

crm.ts

Inline lead capture forms wired to Supabase. Auto-tags leads by page source. Feeds into tenant CRM dashboard.

Reservations

4-State Booking Machine

reservations.ts

1. Party size → 2. Date & time slot → 3. Contact info → 4. Confirmation. Supports Resy, OpenTable, phone, and walk-in flows. Posts to /api/[slug]/reservations.

The Key Insight

RESEARCH FOUNDATION

Built on Proven Best Practices

Best Practice	Why It Matters	Applied How
thinking_level=HIGH	Allocates max tokens for internal reasoning before code generation	Set on every Gemini 3.1 Pro call
temperature=1.0	Lowering temp causes instability and degraded logic in Gemini 3.1 Pro	Never changed — locked at 1.0
Context Caching	90% token reduction for cached design system	DesignSystem JSON cached, reused across all pages
3-Phase Mega-Prompt	Separates architecture, UI/UX, compliance concerns	Phase 1 (HTML), Phase 2 (components), Phase 3 (modules)
Negative Constraints Last	Model retains DO NOT rules through lengthy generation	Always final block in every prompt
Multimodal QA	Model debugs its own visual output via screenshot review	Screenshot → Gemini → CSS patches loop
WCAG 2.2 AA	Mandatory accessibility for all 2026 web builds	aria-labels, 4.5:1 contrast, keyboard nav in every prompt
Schema.org JSON-LD	AEO/GEO: AI crawlers need structured data, not just SEO	Every page includes LocalBusiness or industry-specific schema

EXAMPLE

Italian Restaurant → Live Website

Cucina Verona, Hayes Valley SF — from description to deployed in ~5 minutes

INPUT

name: "Cucina Verona"
industry: restaurant
personality: intimate, rustic-modern
booking: resy
modules: kit, scribe, reservations
competitors: Quince, SPQR

OUTPUT

✓ Design System JSON (cached)
✓ SiteMap: 4 pages, 24 sections
✓ 4 mega-prompts + live data hooks
✓ 4 Next.js page components
✓ design-tokens.ts generated
✓ Live data bound (menu, availability)
✓ Kit + Scribe + CRM + Reservations
✓ QA score: 89% (all pass)
✓ Live on Cloudflare Pages

From Descriptionto Deployed Website.

The Complete Pipeline

Intake

Design System

Site Architecture

Mega-Prompts

Page Generation

Module Injection

QA Agent

Delivery

The Design System — Global Contract

One Mega-Prompt Per Page

Design-System-Native Modules

Kit

Scribe

CRM

Reservations

The Key Insight

Built on Proven Best Practices

Italian Restaurant → Live Website

INPUT

OUTPUT

From Descriptionto Deployed Website.

The Complete Pipeline

Intake

Design System

Site Architecture

Mega-Prompts

Page Generation

Module Injection

QA Agent

Delivery

The Design System — Global Contract

One Mega-Prompt Per Page

Design-System-Native Modules

Kit

Scribe

CRM

Reservations

The Key Insight

Built on Proven Best Practices

Italian Restaurant → Live Website

INPUT

OUTPUT

From Description
to Deployed Website.

From Description
to Deployed Website.