Stage 0: Client Interview

The system presents 5 fast-path questions in sequence:

After Q4, the tier auto-suggestion fires: website = true AND content_copy = true triggers a growth suggestion, which matches Q2 — no change prompt appears.

After Q5, the system asks: "Want to go deeper? A few more questions help us produce better results." For this test, answer y to exercise the extended path.

Extended questions (answer a subset):

The system shows a confirmation summary:

──────────────────────────────────────────
  Vandoko Intake Summary
──────────────────────────────────────────
  URL:          meridianrobotics.com
  Tier:         growth
  Deliverables: Brand Guidelines, Website, Email Templates, Social Content, Content Copy
  Brand status: refresh
  Context:      Series B industrial automation company...
  Timeline:     standard (no hard deadline)
  Inspiration URLs: 3 provided
──────────────────────────────────────────
  [Enter] to confirm and begin pipeline
  Type 'restart' to re-enter answers

Press Enter to confirm.

Verification checklist:

• InterviewResult JSON is assembled with all 7 categories populated
• meta.question_count reflects the actual count of questions answered
• interview_path is "extended"
• deliverable_selections.website is true

Stage 1: INTAKE

The system scrapes meridianrobotics.com via Playwright (client-intake skill, researcher worker, haiku model):

• Homepage, /about, /services — up to 4 pages
• Extracts brand colors (primary, secondary from CSS custom properties)
• Detects typography (heading + body font families)
• Classifies industry using taxonomy: matches manufacturing or b2b-general
• Detects tech stack from HTML markers (WordPress, Next.js, etc.)
• Finds social links (LinkedIn, Twitter/X)
• Computes site quality score (0–100)
• Sets extraction confidence (e.g., 0.85 if Playwright succeeded on 3 pages)

Verification checklist:

• vandoko-pipeline-state.json exists with currentStage: "INTAKE" then updated to "DISCOVERY"
• history array has an entry: { stage: "INTAKE", status: "completed" }
• ClientProfile artifact contains: brand.colors.primary (hex), brand.typography.headings, business.industry (canonical enum), business.tech_stack[], has_social.linkedin
• extraction_confidence is >= 0.5
• StageOutputEnvelope wraps the ClientProfile with stage: "INTAKE", status: "complete"

Stage 2: DISCOVERY

The system conducts research using Claude + Gemini + Exa (discovery-research skill, researcher worker, haiku model):

1. Exa searches for 3–5 competitors (Universal Robots, FANUC will appear since we named them)
2. Claude synthesizes market landscape: trends, opportunities, threats
3. Gemini researches design trends for manufacturing/industrial websites
4. Claude compares Meridian's current brand against competitors (brand gap analysis)
5. Exa discovers 5–8 inspiration references (best manufacturing website designs)
6. Gemini visually analyzes top 3–5 inspiration URLs (extracts color palettes, typography, layout patterns)
7. Exa researches Meridian's brand perception (reviews, press mentions)
8. Context7 looks up framework docs if tech stack was detected

Verification checklist:

• DiscoveryReport artifact contains: competitors[] (3–5 entries), market_landscape.trends[] (3+ items), design_trends.industry_specific[] (3+ items), brand_gaps (positioning_gap, visual_gap, messaging_gap as non-null strings)
• inspiration[] has 5–8 entries with relevance_score values
• visual_references[] has entries with color_palette_extracted hex arrays
• generated_by is "combined" (Claude + Gemini)
• Pipeline state history shows DISCOVERY as completed

Stage 3: STRATEGY

Claude crafts a complete brand strategy (brand-strategy skill, strategist worker, opus model). No external tools — pure synthesis.

The system produces and presents a Strategy Summary Card:

Brand Direction: Meridian Robotics

Positioning Statement
> For VP Operations and Plant Managers at mid-market manufacturers,
> Meridian Robotics is the automation partner that bridges the gap
> between complex robotic technology and practical factory-floor results.

Value Proposition
Precision automation that works on day one -- no PhD required.

Tagline Candidates
1. Precision at Work
2. Built for the Factory Floor
3. Automation, Simplified

Color Palette
- Primary: #1E3A5F — communicates industrial trust and engineering precision
- Secondary: #4A6FA5 — lighter steel blue for depth
- Accent: #FF6B35 — safety orange signals action and energy
- Neutrals: #0D1117 (background), #1A2332 (card), #E8ECF0 (text), #8B95A3 (muted)

Typography
- Display: Space Grotesk, 700 — geometric, technical, modern
- Body: Inter, 400 — clean readability

Template Recommendation
streamline — optimized for logistics and manufacturing verticals

Tone of Voice
professional-warm — short declarative sentences, technical but accessible

Verification checklist:

• BrandStrategy artifact has: positioning.statement (filled template), visual_direction.color_system (all hex, no oklch), visual_direction.template_recommendation (valid slug), tone_of_voice.em_dash_policy = "never"
• voice_spectrum has 5 integer scores (1–5 each)
• decisions[] in pipeline state has: { stage: "STRATEGY", selected: "approve" }
• Lineage entry recorded with creative_selections[]

Stage 4: CONCEPT

Gemini/Nanobanana generates 3 visual concepts (growth tier) via the concept-gallery skill, concept-maker worker, sonnet model:

1. Meridian — Clean (template: streamline) — Hero, cards, buttons images
2. Meridian — Bold (template: plasma) — Hero, cards, buttons images
3. Meridian — Authentic (template: relative) — Hero, cards, buttons images

Plus shared assets: background pattern + icon set (2 more Nanobanana calls). Plus per-concept: typography specimen, gradient options, icon grid. Plus voice signal extraction (text-based radar).

Total Nanobanana calls: ~15–20.

The system builds {output_root}/concepts/gallery.html — a single self-contained HTML file with inline CSS showing all concepts side-by-side. Concept enhancement runs: mood boards, palette visualizations, inspiration lineage added to each concept.

Verification checklist:

• {output_root}/concepts/ directory contains: concept-clean/, concept-bold/, concept-authentic/, shared/
• Each concept directory has: hero.png, cards.png, buttons.png
• gallery.html opens in a browser and shows 3 cards with images, color swatches, and template badges
• ConceptSelection artifact: selected: "concept-bold", selectedAt is ISO-8601
• BrandReference assembled at {output_root}/brand-reference.json

Stages 5–8: Parallel Execution

After concept selection, the pipeline fans out across BRAND, SOCIAL, WEBSITE, and CONTENT simultaneously:

CONCEPT (approved) ──┬── BRAND ──── EMAIL ──┐
                     ├── SOCIAL             ├── LAUNCH
                     ├── WEBSITE ── QA ─────┤
                     └── CONTENT ───────────┘

BRAND (brand-guidelines skill, brand-builder worker, haiku): Generates brand guidelines PPTX (10+ slides via PptxGenJS), design tokens directory.

SOCIAL (social-pack skill, content-creator worker, sonnet): Produces platform banners for LinkedIn, Facebook, Instagram, Twitter (growth tier). Post templates PPTX, content calendar markdown.

WEBSITE (website-build skill, site-builder worker, sonnet): Translates BrandStrategy + ConceptSelection into ClientBrief format, delegates to website-autopilot. Installs vandoko-theme if vandoko-branded. Resolves component registry. Generates design token CSS. Captures e2e baseline screenshots. Scaffolds deliverables portal.

CONTENT (content stage, via ben-hamine/copywriter): SEO pages, GEO content, blog posts, articles, whitepapers, lead magnets.

When BRAND completes: EMAIL (email-pack skill, content-creator worker, sonnet) launches. Produces welcome + follow-up sequences (2 sequences, 2 personas for growth tier) in Zoho + HubSpot ESP formats.

Verification checklist:

• Pipeline state history shows BRAND, SOCIAL, WEBSITE, CONTENT all started before any completed
• Each parallel stage has its own history entry with timestamps
• BRAND: {output_root}/brand/ contains PPTX file + tokens directory
• SOCIAL: {output_root}/social/ contains banner images + templates
• WEBSITE: {output_root}/website/ contains built site, deployUrl in artifacts
• EMAIL: {output_root}/email/ contains HTML email files + ESP import instructions
• CONTENT: contentPack artifact has seoPages, blogPosts, articles counts

Stage 9: QA

The QA orchestrator (cross-team call to marketing-website-qa) runs 7 audit dimensions against the built website:

Produces a QualityReport with overall_score, grade (A–F), pass (boolean), and top 10 issues sorted by severity.

Verification checklist:

• QualityReport exists with overall_score (0–100), grade, pass
• Each dimension has a sub-score
• Top issues array contains up to 10 items with location and fix fields
• If score >= 80: pass = true, proceed to LAUNCH
• If score < 80: launch-prep halts, recommends site-polish

Stage 10: LAUNCH

The launch-prep skill (launch-gatekeeper worker, sonnet model):

1. Delegates QA to qa-orchestrator (if not already run)
2. Compiles handoff document at {output_root}/launch/handoff.md
3. Prepares deployment commands (does NOT execute them)
4. Computes lineage scoring aggregates

Verification checklist:

• handoff.md exists with all sections populated: Pipeline Summary, Decision History, Deliverables (Brand, Social, Website, Email, Content), Quality, Maintenance Notes
• Pipeline state status = "completed"
• completedAt is ISO-8601
• project-lineage.json has scoring object with: tooling_performance, anti_pattern_score, revision_rate
• Deployment commands are printed but NOT executed