Inboxmate Demo

InboxMate Demo Setup Pipeline

🚨 NON-NEGOTIABLE CHECKLIST (read before building)

A demo is only "done" when ALL of these are true. Skipping any one of them kills conversion and has caused real bugs in past batches:

• Primary brand color fetched via OpenBrand and set on the agent (update_widget_style → primaryColor). Default InboxMate emerald on a customer's demo is the single most visible "we didn't tailor this" signal — never ship without trying OpenBrand. If OpenBrand returns no usable hex (only white/black/near-white/near-black), fall back to manual CSS/HTML inspection; if that also yields nothing, leave default but flag it in the report.
• Logo fetched via OpenBrand and set on the demo page (update_demo_page → logoUrl). Prefer non-favicon entries, pick the largest resolution. A real logo in the demo's company card builds immediate trust; the fallback initials look generic.
• demoVersion: V3 on the CRM opportunity AND version: "v3" on the demo page. Never ship v1 or v2.
• Quick questions are cards, not chips. You MUST call update_quick_questions with style: "cards" and objects in the form { text, title, description, icon }. quick_setup_demo does NOT set these — it's a known pitfall.
• customMessage is sales copy, not agent-greeting copy. The demo page already shows a big "Hi, {Company}! 👋" headline — customMessage is the sub-line under it. Write ABOUT the bot, not AS the bot.
  • ✅ "Wir haben einen KI-Assistenten speziell für [Company] gebaut — trainiert auf [concrete thing]. Testen Sie ihn rechts unten."
  • ❌ "Willkommen bei [Company]! Ich helfe dir gerne weiter…" (that's the agent's greeting — don't duplicate it here)
• useCases is an array of OBJECTS, never strings. Each entry MUST be { "text": "…", "icon": "lucide-name" }. Plain strings like "Beratung anbieten [icon:home]" are a known failure mode — the frontend will render [icon:home] as literal text and fall back to a checkmark glyph. If your first impulse is to stuff the icon into the text, stop and use the object form.
  • ✅ { "text": "Immobilienbewertung automatisch einleiten — …", "icon": "home" }
  • ❌ "Immobilienbewertung automatisch einleiten — … [icon:home]"
• Custom DE greeting set on the agent via update_widget_style → greetingMessageDe. Do NOT rely on the auto-generated "Hallo! Ich bin der {Company}-Assistent. Wie kann ich Ihnen heute helfen?" default — that's the single most generic-feeling line and every prospect sees it. Write one line that references 2-3 concrete topics the bot can answer ("Hallo! Ich beantworte Fragen zu Radreisen, Gruppenreisen und Beratung"). Same rule for greetingMessage (EN) if the site is English.
• Quick questions in the RIGHT language slot. The update_quick_questions tool has THREE fields: questions (bare/locale-agnostic), questionsEn, questionsDe. For a German site you MUST use questionsDe. Subagents have mis-routed cards into the EN slot repeatedly; the widget is locale-aware and will render nothing if DE is empty but EN has cards.
• Knowledge bucket has real site content. Call scrape_and_build_knowledge with 2–3 relevant URLs from the prospect's site (beyond the homepage excerpt). Skip 404s silently; don't abort the demo.
• Agent is republished after any config change (e.g. after update_quick_questions). Skip this and the widget bundle still serves stale config.
• CRM opportunity has campaignId set to the active campaign UUID. Orphan opportunities never get drafts.

If ANY of the above is missing, the demo is not done — loop back and fix before moving on.

Announce to the user at the very start:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
InboxMate Demo Pipeline started.
Working through 6 phases — I'll narrate each step.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

PHASE 0 — Learn the Platform

Announce: [0/5] Learning InboxMate platform capabilities...

You are new to InboxMate. Do this before anything else.

InboxMate is a white-label AI chatbot platform built by psquared. Businesses embed a chat widget on their website — visitors chat with an AI agent that knows the company's products, services, and FAQs.

The MCP server at https://app.psquared.dev/api/mcp exposes tools to create and configure these agents programmatically. All operations run against a shared demo account.

Call tools/list on the MCP server first to get the current tool list and confirm connectivity:

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}
{"jsonrpc":"2.0","id":2,"method":"notifications/initialized","params":{}}
{"jsonrpc":"2.0","id":3,"method":"tools/list","params":{}}

Environment variables (read from .env in the current working directory at startup):

• NUXT_MCP_DEMO_TOKEN — Bearer token for the InboxMate MCP server
• OPENBRAND_API_KEY — API key for OpenBrand color/logo extraction (used in Phase 2f)
• PSQUARED_CRM_TOKEN — Bearer token for the psquared CRM API (used in Phase 5)

MCP connection:

• URL: https://app.psquared.dev/api/mcp
• Auth: Authorization: Bearer <NUXT_MCP_DEMO_TOKEN>
• Transport: JSON-RPC 2.0 over HTTP POST

Key concepts:

• Agent — the AI chatbot entity. Has a system prompt, greeting, widget config (color, domain whitelist, predefined questions)
• Knowledge bucket — a vector store. Contains multiple knowledge items that the agent retrieves at query time
• Knowledge items — individual focused text chunks added to a bucket. Each item should cover ONE topic (overview, pricing, FAQ, contact, etc.)
• whitelistedDomains — controls where the widget appears. Defaults to demo.inboxmate.psquared.dev

After confirming tools list, announce: [0/5] Platform ready. Proceeding to prospect research.

PHASE 1 — Research the Prospect

Announce: [1/5] Researching [Company Name]...

1a — Validate Website First

Before scraping, check if the website is reachable and current. Use WebFetch on the homepage.

Auto-skip if ANY of these are true:

• Website returns HTTP error / timeout / unreachable
• Domain is parked, expired, or "coming soon"
• Page has no meaningful content (just a logo or placeholder)
• Copyright year is 2+ years behind current year — site is abandoned
• Only a social media profile exists (no real website)

If the website is unusable, announce: SKIP: [Company] — [reason]. Website is not suitable for a demo. Do NOT ask the user what to do. Just skip and move on. If running as part of the batch pipeline, the batch skill handles CRM marking. If running standalone, just stop and report the skip reason — the user can decide later.

1b — Scrape Content

Given a company name and/or domain, fetch all of these pages (adjust paths as needed):

Use WebFetch on each URL. Do not skip pages — thin knowledge = bad demo.

Extract and record:

• Company name (exact spelling, including GmbH/AG/Ltd if present)
• Core product/service in one sentence
• Target customer segment
• Top 3 USPs (what makes them different)
• Pricing structure (free trial? subscription? one-time?)
• Primary CTA (book demo / sign up / contact / request quote)
• Contact details (email, phone, address, hours)
• Tone of the website (formal/informal, corporate/friendly)
• Primary language of the website content

After fetching, announce: [1/5] Research complete. Analyzing language and planning content...

PHASE 2 — Language Detection & Content Planning

Announce: [2/5] Planning agent content and language configuration...

2a — Detect Language

Look at the website content you scraped:

Announce your decision: Language: [EN / DE / MULTI] — [reason in one sentence]

2b — Plan Knowledge Items

Do NOT put all content into one blob. Create separate focused knowledge items per topic. Each item is independently retrieved by vector search — short, specific items win over large walls of text.

Plan these items (create all that have content):

Write all item content in the target language(s). For multi: write each item in the language that's most natural for that content, or bilingual if the company operates in both.

2c — Craft the System Prompt

Write a crisp, specific system prompt — not a generic template. Include ALL of these:

You are [Persona Name], the AI assistant for [Company Name].

**Who you help:** [Target customer segment in 1 sentence]

**Your personality and tone:** [Derived from website tone — e.g. "friendly and approachable, using informal language (du/you)" OR "professional and precise, using formal language (Sie/formal English)"]

**Language:** [e.g. "Always respond in German (du-Form)" OR "Respond in the language the visitor uses — German or English"]

**Your goal:** Help visitors understand [Company]'s [main product/service] and guide them toward [primary CTA — e.g. "booking a free demo", "starting a free trial", "contacting the sales team"].

**What you know:** You have access to [Company]'s complete product information, pricing, FAQ, and contact details. Answer from this knowledge.

**When you can't answer:** If a question falls outside your knowledge, say so honestly and offer to connect them with the team: [contact email or "via the contact form on [domain]"].

**Never:** Invent pricing, make promises not reflected in company materials, or discuss competitors in detail.

Adjust heavily based on actual company context. This prompt should sound like it was written for THIS specific company.

2d — Greeting Message

Write in target language. Not generic. Reference the product, their situation, or a hook.

• Bad: "Hi! How can I help you?"
• Good (EN): "Hi! I'm [Name], [Company]'s AI assistant. Ask me anything about our [product] — pricing, features, or how to get started."
• Good (DE): "Hallo! Ich bin [Name], der KI-Assistent von [Company]. Stell mir gerne Fragen zu unseren [Dienstleistungen] — ich helfe dir weiter!"

For multi: write separate EN and DE greetings.

2e — Predefined Quick Questions (Card Tiles)

ALWAYS use card tiles — never plain text pills. Cards look dramatically better and convert higher. They show an icon, a bold title, and a short description — making suggestions feel like real features instead of generic prompts.

4–5 suggestion cards that a real prospect would click. Make them irresistible — they should surface the company's best selling points.

Each question is an object with these fields:

• text: The actual message sent to the chat when clicked. This is what the AI will answer.
• title: Short card title displayed prominently (2-5 words). This is what the user reads first.
• description: Brief hint about what to expect (3-5 words). Shown below the title in smaller text.
• icon: A Lucide icon name in kebab-case (e.g. "briefcase", "credit-card", "shield-check", "clock", "users", "rocket", "globe", "zap", "heart", "target", "lightbulb", "star", "package", "award", "calendar", "mail", "phone", "building", "trending-up", "search"). If invalid, falls back to a checkmark.

Rules:

• Use the target language (both EN + DE arrays for multi)
• Mix types: feature question, pricing question, differentiator question, use-case question
• Keep titles short (2-5 words), descriptions short (3-5 words)
• The text field can be longer — it's the actual question the AI answers
• Always set style: 'cards' in the update_quick_questions call

Examples for a SaaS company:

// EN
[
  { "text": "What does [Product] do and who is it for?", "title": "What is [Product]?", "description": "Features & use cases", "icon": "rocket" },
  { "text": "How much does [Product] cost? Are there different plans?", "title": "Pricing & Plans", "description": "Tiers & free trial", "icon": "credit-card" },
  { "text": "How long does setup take and do I need technical knowledge?", "title": "Getting Started", "description": "Setup in minutes", "icon": "zap" },
  { "text": "What makes [Product] different from competitors?", "title": "Why [Company]?", "description": "Key differentiators", "icon": "award" }
]
// DE
[
  { "text": "Was macht [Produkt] und für wen ist es gedacht?", "title": "Was ist [Produkt]?", "description": "Funktionen & Einsatz", "icon": "rocket" },
  { "text": "Was kostet [Produkt]? Gibt es verschiedene Pakete?", "title": "Preise & Pakete", "description": "Tarife & Testphase", "icon": "credit-card" },
  { "text": "Wie lange dauert die Einrichtung und brauche ich technisches Wissen?", "title": "Erste Schritte", "description": "Setup in Minuten", "icon": "zap" },
  { "text": "Was unterscheidet [Produkt] von der Konkurrenz?", "title": "Warum [Company]?", "description": "Die Vorteile", "icon": "award" }
]

2f — Brand Color (via OpenBrand API)

Use the OpenBrand API to extract the brand color — it's more reliable than manual CSS inspection.

API call:

curl "https://openbrand.sh/api/extract?url=https://[domain]" \
  -H "Authorization: Bearer $OPENBRAND_API_KEY"

The OPENBRAND_API_KEY is in the .env file (read it at startup along with the other tokens).

Response format:

{
  "success": true,
  "data": {
    "brandName": "Company Name",
    "logos": [{ "url": "...", "type": "favicon", "resolution": {...} }],
    "colors": [
      { "hex": "#3b82f6", "usage": "primary" },
      { "hex": "#64748b", "usage": "secondary" },
      { "hex": "#ffffff", "usage": "accent" }
    ]
  }
}

How to pick the color:

1. Use the color with "usage": "primary" from the response
2. NEVER use pure black (#000000) or pure white (#ffffff) as the widget color — even if it's the brand's primary. These look broken in the widget. If the primary is black or white, use the secondary color instead.
3. If the primary color is too light (very pale pastel) or too dark (near-black), use the next best color from the response
4. If OpenBrand returns an error or no colors, fall back to manual extraction from CSS/HTML
5. If the primary color looks washed out for a widget accent, prefer the darkest non-white, non-black color returned

Also grab the logo URL from data.logos — use the highest-resolution one as logoUrl in the demo page.

Manual fallback is MANDATORY when OpenBrand has no hit. Shipping a prospect's demo with the default InboxMate emerald is the single most obvious "we didn't tailor this" signal — it's not an acceptable default when we can do better in 30 seconds. Cases that commonly need the fallback:

• OpenBrand returns 200 but data.colors is empty or only contains near-white / near-black / pure grayscale (e.g. #000, #111, #fff, #eee).
• OpenBrand times out or the site is temporarily down. Try https://web.archive.org/web/<YEAR>/https://<domain> (use the current year first, then a prior year) to pull the cached version.
• Site passed but OpenBrand misread the brand — e.g. picked a background accent rather than the real brand color.

Fallback procedure:

1. WebFetch the homepage. If it errors or OpenBrand failed, also try archive.org.
2. Ask yourself: what hex is used for the primary CTA / header bar / logo mark? Read the inline CSS or rendered markup.
3. Pick one hex (kebab format like #1a365d). Skip pure grayscale (black/white and anything brightness < 30 or > 230 after weighted-RGB calc).
4. If the site truly has nothing colored (extremely rare — some minimalist funeral/medical sites), pick a thematically appropriate muted color (e.g. deep navy #2c3e50 for professional services, forest green #505d46 for nature/funeral, slate #4a5568 for neutral) rather than leaving emerald.
5. Pass the chosen hex to update_widget_style and republish the agent.

Logo fallback when OpenBrand has no usable logo: grab the <img> used in the site header, or the apple-touch-icon from <link rel="apple-touch-icon">, or an og:image meta tag. Anything real is better than the auto-generated initials badge.

2g — Demo Page Content

Offer deadline — ASK THE USER

Before setting the offer, ask the user for the deadline. Do NOT assume 7 days or any other duration.

Ask:

When should the demo offer expire?
Examples: "in 14 days", "2026-04-01", "end of month"

Wait for the user's answer. Convert their response to an ISO 8601 date for offerExpiresAt.

If running as part of /inboxmate-batch-demo: Ask ONCE at the start for all demos in the batch (e.g., "All demos expire on 2026-04-15"). Do NOT ask per company.

Offer text — STRICT RULES

• offerText: The headline shown above the countdown timer. It describes the limited-time offer the prospect gets if they sign up before the deadline.

FORBIDDEN — never use any of these:

• "Kostenlose Erstberatung" (free consultation)
• "Kostenlose Beratung" (free consultation)
• "Kostenloses Erstgespräch" (free initial call)
• Any variation of "free consultation/call/meeting" — we are NOT offering consultations

What the offer IS: A time-limited discount or special deal for signing up to InboxMate (the chatbot product). The countdown shows when this deal expires.

Use one of these patterns:

• DE: "Jetzt starten und bis zu 50% Rabatt sichern" / "Sonderkonditionen für Ihren KI-Chatbot — nur bis [date]" / "Ihren KI-Assistenten jetzt aktivieren — exklusive Konditionen"

• EN: "Start now and save up to 50% in your first year" / "Special pricing for your AI chatbot — limited time" / "Activate your AI assistant — exclusive terms available"

• customMessage: SALES-PAGE copy, NOT agent greeting. 1–2 sentences speaking ABOUT the bot TO the prospect — never in the bot's own voice. The demo page already shows a big "Hi, {Company}! 👋" headline; customMessage is the sub-headline under it.

  Must hit the Value Equation levers — see ~/.claude/skills/inboxmate-demo-with-outreach/references/value-equation.md. Every customMessage should include at least:

  • a specific thing we built for them (perceived likelihood),

  • an explicit speed signal ("in 10 Minuten live", "läuft bereits"),

  • a removed friction ("kein Setup", "keine Anmeldung").

  • ✅ Right: "Ihr KI-Assistent für [Company] ist fertig — trainiert auf [specific area], in 10 Minuten auf Ihrer Seite live. Einfach rechts unten testen, kein Setup, keine Anmeldung."

  • ⚠️ Only-hook, missing levers: "Wir haben einen KI-Assistenten speziell für [Company] gebaut — trainiert auf Ihre Leistungen. Testen Sie ihn rechts unten." (acceptable but weak — no TTV, no removed friction)

  • ❌ Wrong: "Willkommen bei [Company]! Ich helfe dir gerne weiter — ob X, Y oder Z." (agent greeting — not for this field)

After planning, show a summary to the user:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
PLAN READY — confirm before building:
  Company: [name]
  Language: [en/de/multi]
  Knowledge items: [N items listed by title]
  Quick questions: [list]
  Color: [hex]
  Offer: [offerText]
  Deadline: [offerExpiresAt]
Proceeding to build in 5 seconds unless you say stop.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

PHASE 3 — Build the Agent via MCP

Announce: [3/5] Building agent in InboxMate...

Use individual MCP tools — do NOT use quick_setup_demo for the full pipeline. It only supports one knowledge item. We need multiple.

Step 3.1 — Create Agent

{
  "method": "tools/call",
  "params": {
    "name": "create_agent",
    "arguments": {
      "name": "[Company] Assistant",
      "prompt": "[full system prompt from 2c]",
      "primaryColor": "[hex from 2f]",
      "greetingMessage": "[EN greeting from 2d — omit if DE-only]",
      "greetingMessageDe": "[DE greeting from 2d — omit if EN-only]",
      "buttonIcon": "[choose based on company personality — see guide below]",
      "buttonShape": "[circle, squared, or icon-only — see guide below]",
      "agentIconType": "[choose based on company personality — see guide below]",
      "widgetPresence": "[shimmer or calm — see guide below]",
      "uiLang": "[en, de, or multi — must match language from Phase 2]"
    }
  }
}

Widget appearance guide — ALWAYS set these fields:

Rules:

• buttonIcon and agentIconType MUST be different from each other
• Never default to robot — choose based on the company's personality and industry
• When in doubt: buttonIcon: "messageCircle", agentIconType: "inboxmate", buttonShape: "circle", widgetPresence: "calm"

Save agentId from the response. Announce: Agent created: [agentId]

Step 3.2 — Create Knowledge Bucket

{
  "method": "tools/call",
  "params": {
    "name": "create_knowledge_bucket",
    "arguments": {
      "name": "[Company] Knowledge",
      "description": "Website knowledge base for [Company] — [N] items"
    }
  }
}

Save bucketId. Announce: Knowledge bucket created: [bucketId]

Step 3.3 — Add Knowledge Items

For each item planned in 2b, call add_to_bucket separately.

CRITICAL: Always include sourceUrl. This enables source citations in the chatbot — the wow effect for prospects. The URL must point to the actual page the content was scraped from (e.g. https://company.at/leistungen for the services item, https://company.at/kontakt for the contact item). If content came from the homepage, use the homepage URL. Never omit sourceUrl.

{
  "method": "tools/call",
  "params": {
    "name": "add_to_bucket",
    "arguments": {
      "bucketId": "[bucketId]",
      "title": "[item title]",
      "content": "[item content — focused, complete, in target language]",
      "sourceUrl": "https://[domain]/[exact page path this content came from]"
    }
  }
}

Announce after each: Knowledge item added: "[title]" (source: [url])

Wait for each call to complete before the next — do NOT batch these.

Faster alternative: Instead of adding items one by one, use the scrape_and_build_knowledge MCP tool which scrapes multiple URLs in one call and creates knowledge items with real sourceUrls automatically:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "scrape_and_build_knowledge",
    "arguments": {
      "bucketId": "[bucketId]",
      "urls": ["url1", "url2", "url3", ...],
      "agentId": "[agentId]"
    }
  }
}

This is recommended for 5+ URLs. It handles scraping, chunking, and embedding in one operation.

Step 3.3b — (If multi-lang) Add Language Support item

If language is multi: add one extra knowledge item to the bucket:

• Title: "Sprachunterstützung / Language Support"
• Content:

  This AI assistant supports both German and English. It responds in whatever language the visitor uses.
  
  Key features relevant for [Company]:
  - Available 24/7, responds instantly in German or English
  - Trained on [Company]'s complete product and service information
  - Handles common customer questions so your team doesn't have to
  - Guides visitors toward [primary CTA]
  - Escalates to your team for complex inquiries
  
  Use cases for [Company]:
  [Write 2-3 specific use cases based on what you know about the company]
  

Content quality rules:

• Each item: 200–800 words. Long enough to be useful, short enough for precise retrieval.
• Write in complete sentences, not bullet-point dumps
• Include actual numbers, prices, and specifics from the website
• For FAQ items: include both the question AND the answer, verbatim if possible

Step 3.4 — Link Bucket to Agent

{
  "method": "tools/call",
  "params": {
    "name": "set_knowledge",
    "arguments": {
      "agentId": "[agentId]",
      "knowledgeBucketIds": ["[bucketId]"]
    }
  }
}

Step 3.5 — Set Quick Questions (Card Format)

{
  "method": "tools/call",
  "params": {
    "name": "update_quick_questions",
    "arguments": {
      "agentId": "[agentId]",
      "style": "cards",
      "questionsEn": [
        { "text": "[message sent to chat]", "title": "[Card Title]", "description": "[3-5 word hint]", "icon": "[lucide-icon-name]" },
        { "text": "[message sent to chat]", "title": "[Card Title]", "description": "[3-5 word hint]", "icon": "[lucide-icon-name]" }
      ],
      "questionsDe": [
        { "text": "[Nachricht an den Chat]", "title": "[Karten-Titel]", "description": "[3-5 Wort Hinweis]", "icon": "[lucide-icon-name]" },
        { "text": "[Nachricht an den Chat]", "title": "[Karten-Titel]", "description": "[3-5 Wort Hinweis]", "icon": "[lucide-icon-name]" }
      ]
    }
  }
}

For EN-only: use questionsEn only, leave questionsDe as []. For DE-only: use questionsDe only, leave questionsEn as []. Always set style: "cards" — this enables the rich card display in the widget.

Step 3.5b — Restrict widget to demo page only

After the demo page is created (Phase 4), the MCP automatically calls update_widget_style to set showOnPages: ['/?id=<demoId>']. This is handled by the create_demo_page tool. You do NOT need to call update_widget_style manually for this.

Step 3.6 — Publish Agent

{
  "method": "tools/call",
  "params": {
    "name": "publish_agent",
    "arguments": { "agentId": "[agentId]" }
  }
}

Announce: Agent published and live.

PHASE 4 — Create Demo Page

Announce: [4/5] Creating demo page...

{
  "method": "tools/call",
  "params": {
    "name": "create_demo_page",
    "arguments": {
      "agentId": "[agentId]",
      "companyName": "[Company Name]",
      "companyDomain": "[domain.com]",
      "logoUrl": "[logo URL if found — otherwise omit]",
      "offerText": "[from 2g]",
      "offerExpiresAt": "[ISO date 7 days from today]",
      "customMessage": "[from 2g]",
      "language": "[en or de — must match the language you chose in Phase 2]",
      "version": "v3",
      "useCases": [
        { "text": "[Use case 1 — specific to this company, 1 sentence]", "icon": "[lucide-icon-name]" },
        { "text": "[Use case 2 — specific to this company, 1 sentence]", "icon": "[lucide-icon-name]" },
        { "text": "[Use case 3 — specific to this company, 1 sentence]", "icon": "[lucide-icon-name]" },
        { "text": "[Use case 4 — optional, if relevant]", "icon": "[lucide-icon-name]" }
      ]
    }
  }
}

The MCP will automatically restrict the widget to only show on /?id=<demoId> — no extra step needed.

Save demoId and playgroundUrl.

PHASE 5 — Create Opportunity in CRM

Announce: [5/6] Creating CRM opportunity...

Create an opportunity in the CRM so the demo enters the review pipeline:

curl -s -X POST https://crm.psquared.dev/graphql \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $PSQUARED_CRM_TOKEN" \
  -d "{\"query\":\"mutation { createOpportunity(data: { name: \\\"[Company Name] — InboxMate Demo\\\", stage: SCREENING, demoStatus: PENDING_REVIEW, demoType: CHATBOT, demoUrl: { primaryLinkUrl: \\\"[playgroundUrl]\\\" }, companyId: \\\"[companyId]\\\", outreachType: INBOXMATE, demoVersion: V3 }) { id name stage demoStatus } }\"}"

> **`demoType: CHATBOT` is required** — it separates chatbot demos from inbox demos (`/inboxmate-inbox-demo`, `demoType: INBOX`) in campaigns, drafts and CRM views. Opportunities with `demoType: null` are treated as legacy CHATBOT.

Note: The companyId comes from the CRM. If this company doesn't exist in the CRM yet (standalone demo, not from batch pipeline), look it up first or create it.

Announce: Opportunity created at SCREENING / PENDING_REVIEW — ready for /review-demos

PHASE 6 — Deliver

Announce: [6/6] Done!

Output this summary to the user:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
DEMO READY — [Company Name]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Playground URL (share this with the prospect):
→ [playgroundUrl]

What was configured:
  Language:        [en / de / multi]
  Brand color:     [hex]
  System prompt:   [first 2 sentences of the prompt]
  Greeting (EN):   [greeting]
  Greeting (DE):   [greeting or "—"]
  Quick questions: [list, language-appropriate]
  Knowledge items: [N items — list titles]
  Offer:           [offerText] (expires [date])

Agent ID: [agentId] (for follow-up adjustments)
Demo ID:  [demoId]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Follow-Up Adjustments

If the demo needs changes after delivery, use these individual tools:

Quality Gate — Before Delivering

Run through this checklist mentally before Phase 5:

• System prompt is specific to this company — no generic filler
• Greeting message references the product or company name
• Quick questions use card format (objects with text, title, description, icon) — NOT plain strings
• Quick questions would make a real prospect click them
• Primary color matches the company brand — NOT pure black (#000000) or pure white (#ffffff)
• At least 4 knowledge items covering: overview, services, pricing/FAQ, contact
• Every knowledge item has a sourceUrl — the exact page URL it was scraped from (enables source citations in chat)
• If multi-lang: DE and EN questions both filled, greeting in both languages
• If multi-lang: knowledge item 7 (language support + use cases) added
• Knowledge items are focused topics, not one big dump
• Offer text is about InboxMate pricing/discount — NOT "Kostenlose Beratung/Erstberatung/Erstgespräch"
• buttonIcon is set and matches company personality — NOT defaulting to robot
• agentIconType is set and DIFFERS from buttonIcon
• buttonShape and widgetPresence are set
• Widget domain is restricted to demo.inboxmate.psquared.dev (auto-set by the platform)
• language matches the company's website language
• 3-4 specific useCases as objects with { text, icon } — not plain strings
• Use case icons are valid Lucide kebab-case names (e.g. "clock", "users", "shield-check")
• Offer deadline was confirmed by the user (not randomly picked)
• Offer text does NOT contain "Beratung", "Erstberatung", "Erstgespräch", or any consultation language
• Widget domain + page restriction set (auto-handled by create_demo_page)