This document defines 5 diverse running examples that appear throughout the book to demonstrate concepts in context. Each example is designed to highlight different aspects of the Synergy Triangle framework.
Focus: Vibe Coding Type: Startup (2-person team) Domain: Logistics/Operations AI Complexity: Medium (LLM + retrieval + agents)
QuickShip is a small logistics startup helping small e-commerce businesses optimize their shipping routes and carrier selection. The founding team consists of a developer and a former logistics manager. They use vibe coding to rapidly prototype and iterate on their product.
| Chapter | Example Content | Callout Type |
|---|---|---|
| 1 - AI Capabilities | QuickShip decides to build an AI-powered route optimizer. Introduction to the problem space. | big-picture: Why logistics is ripe for AI disruption |
| 2 - Synergy Triangle | How QuickShip's developer handles vibe coding while the logistics manager provides domain expertise. | key-insight: The synergy between domain expert and vibe coder |
| 3 - Understanding Vibe Coding | QuickShip's developer shows their first vibe coding session, using Cursor to build a route visualization dashboard | practical-example: First prompt to working UI in 20 minutes |
| 4 - Vibe Coding Workflows | Iterating on the route algorithm. Multi-file generation for the full dashboard | code: Sample prompts that generate complex multi-file structures |
| 5 - Full Lifecycle | Moving from prototype to testable code. Adding unit tests with AI help | skill: Using Claude Code to generate test cases |
| 6 - Skills & Plugins | QuickShip creates a custom skill for their deployment workflow | document: SKILL.md for their deployment pipeline |
| 7 - AI PM Strategy | The logistics manager defines success metrics: "95% of route suggestions must be within 10% of optimal" | document: PRD excerpt with eval criteria |
| 8 - AI PM Toolkit | Building eval datasets from historical shipping data | code: Python script for creating golden datasets |
| 9 - Discovery & Prioritization | User research with small e-commerce owners | team-discussion: Conversation between founder and users |
| 11 - Ethics & Trust | Addressing bias in route recommendations (some areas have fewer carrier options) | architecture: Fallback logic for underserved regions |
| 12 - AI Engineering Principles | Transitioning from vibe prototype to production architecture | architecture: System diagram showing the production stack |
| 13 - Prompts & Agents | Implementing a multi-agent system: one agent for route calculation, one for carrier selection | code: LangGraph agent implementation |
| 15 - Model Selection | Router strategy: using lightweight models for simple routes, frontier models for complex multi-stop routes | document: Model routing decision matrix |
| 16 - Inference & LLMOps | Cost monitoring dashboard. Optimizing token usage | code: Cost tracking implementation |
| 17 - Scaling & Observability | Monitoring latency for real-time route suggestions | architecture: Observability stack with Prometheus/Grafana |
| 18 - Integration Playbook | QuickShip's complete vibe-to-production journey | big-picture: Timeline from idea to launch |
| 19 - Case Studies | Lessons learned: what worked, what failed | key-insight: The importance of eval-first development |
Code Callout (Ch 4):
# QuickShip prompt: Generate route visualization dashboard
"Create a React dashboard showing delivery routes on a map.
Include a sidebar with route details, carrier info, and cost
breakdown. Use Tailwind CSS. Make it interactive with hover
states showing delivery windows."
Skill Callout (Ch 6):
# SKILL.md for QuickShip Deployment
## Triggers
- "deploy to staging"
- "prepare release"
## Instructions
1. Run full test suite
2. Build Docker image
3. Deploy to staging cluster
4. Run smoke tests
5. Post status to #quickhip-slack
Team Discussion (Ch 9):
PM (former logistics manager): "Users keep asking for
real-time tracking. But they also complain about subscription
costs. We need to understand if tracking is a differentiator
or just noise."
Developer: "I can add basic tracking with webhooks from carriers.
Won't need new AI features. Just data aggregation."
Focus: AI PM Type: Healthcare SaaS (mid-size company) Domain: Healthcare analytics AI Complexity: High (LLM + RAG + strict evals + compliance)
HealthMetrics is a B2B platform helping hospital administrators analyze patient flow, resource utilization, and staff scheduling. Their AI features include natural language queries of operational data and automated report generation.
| Chapter | Example Content | Callout Type |
|---|---|---|
| 1 - AI Capabilities | HealthMetrics evaluates AI for clinical decision support | warning: Regulatory constraints on healthcare AI |
| 2 - Synergy Triangle | PM team works with compliance officers and clinical advisors | key-insight: Healthcare requires extra PM rigor |
| 6 - AI PM Strategy | Build/buy/bake decision: building custom RAG on their medical literature | document: Build vs buy comparison table |
| 7 - AI PM Strategy | Categorizing their product: copilot (query interface) vs autonomous (auto-scheduling) | big-picture: AI product taxonomy in healthcare |
| 8 - AI PM Toolkit | The USID.O framework adapted for healthcare: longer compliance checkpoints | document: Modified USID.O checklist |
| 8 - Eval-First PRDs | Defining "safe" metrics: AI suggestions must never conflict with clinical guidelines | document: PRD excerpt with hard constraints |
| 9 - User Research | Interviewing hospital administrators about pain points | team-discussion: Research synthesis session |
| 10 - Ethics & Governance | ISO 42001 compliance, bias detection in patient routing | document: Compliance checklist |
| 10 - Bias Detection | Finding that the AI deprioritizes patients from lower-income areas | warning: Real-world bias discovered in eval |
| 11 - GTM | Pricing model: per-query vs per-bed vs enterprise flat fee | document: Pricing model comparison |
| 12 - AI Engineering Principles | Eval-driven development with clinical safety as priority | key-insight: Why healthcare can't skip evals |
| 13 - RAG Systems | Building medical literature RAG with strict relevance scoring | architecture: RAG pipeline with guardrails |
| 14 - Advanced RAG | Graph RAG for connecting patient symptoms to resource availability | code: Neo4j integration for knowledge graphs |
| 15 - Model Selection | Router using Claude for complex queries, smaller models for routine reports | document: Model selection matrix with latency/cost/accuracy |
| 16 - Guardrails | Input validation preventing PII from entering the prompt | code: PII detection and redaction module |
| 17 - Observability | Monitoring "escalation rate" - how often AI defers to human judgment | big-picture: Healthcare-specific AI metrics |
| 18 - Integration Playbook | HealthMetrics' 18-month journey from concept to enterprise rollout | key-insight: Healthcare AI timelines |
| 20 - Team Structures | Specialized roles: Clinical AI PM, Compliance AI Lead, AI Safety Engineer | document: Org chart for healthcare AI team |
PRD Excerpt (Ch 8):
## Success Metrics
- Query accuracy: >= 90% on clinical terminology test set
- Response latency: p99 < 3 seconds
- Safety escalation rate: <= 5% (AI defers to human)
- Zero hallucination tolerance on medication interactions
## Hard Constraints
- AI must never suggest overriding physician orders
- All patient data must be anonymized in prompts
- Audit trail required for all AI recommendations
Bias Warning (Ch 10):
## Bias Detection Finding
During eval on 10,000 historical patient records, we discovered:
- AI deprioritized patients from zip codes with median income < $45k
- False negative rate 3x higher for patients with non-commercial insurance
- Required adding fairness constraints to the model router
Team Discussion (Ch 9):
Senior PM: "Dr. Martinez says the query interface works great
for simple questions. But she never uses it for complex cases."
Junior PM: "Is that because she doesn't trust the AI for complex
cases, or because the interface is too slow?"
Senior PM: "She wants to verify the source. Show me the evidence.
That's our next sprint: citation links and confidence scores."
Focus: AI Engineering Type: Enterprise software company Domain: Data pipeline/ETL automation AI Complexity: Very High (multi-agent orchestration + RAG + LLMOps)
DataForge helps enterprises automate their data pipelines using natural language. Users describe what data transformation they want in plain English, and DataForge's AI generates and maintains the corresponding ETL code.
| Chapter | Example Content | Callout Type |
|---|---|---|
| 2 - Synergy Triangle | DataForge's engineering team needs to understand PM requirements before building | key-insight: Why engineering can't ignore PM |
| 6 - Skills | Engineering team creates shared skills for their ML pipeline development | skill: Multi-agent orchestration skill for data validation |
| 12 - AI Engineering Principles | Eval-driven development: building evals before writing code | big-picture: Enterprise AI development differs from startups |
| 13 - Prompts & Agents | Multi-agent architecture: Parser Agent, Code Generator Agent, Validator Agent | architecture: Agent orchestration diagram |
| 13 - MCP Connectors | Integrating with enterprise data sources via MCP | code: MCP server implementation for Snowflake |
| 14 - Advanced RAG | RAG on internal enterprise data schemas and documentation | code: Schema-aware retrieval implementation |
| 15 - Model Selection | Dynamic model selection based on query complexity | document: Router configuration with thresholds |
| 16 - Inference Serving | vLLM deployment for low-latency code generation | architecture: Inference stack with batching |
| 16 - CI/CD for Prompts | Prompt version control with A/B testing | code: Git-based prompt deployment workflow |
| 17 - Scaling | Multi-tenancy with enterprise isolation | architecture: Tenant isolation architecture |
| 17 - Observability | Custom metrics: code acceptance rate, pipeline success rate, user satisfaction | code: Prometheus metrics definitions |
| 18 - Integration Playbook | DataForge's 2-year journey to production | big-picture: Enterprise sales cycles |
| 19 - Case Studies | Lessons from building in the enterprise | key-insight: Why enterprise AI is slower but more stable |
| 20 - Team Structures | Roles: AI Pipeline Engineer, Prompt Engineer, ML Ops, Enterprise Integration Lead | document: Team structure diagram |
| 21 - Future-Proofing | Plans for autonomous pipeline maintenance agents | research-frontier: Agentic AI for enterprise |
Architecture Diagram (Ch 13):
User: "Create a pipeline that joins customer data with orders,
filters for high-value customers, and outputs to Redshift"
Agents:
[Parser Agent] -> [Code Generator Agent] -> [Validator Agent]
| | |
v v v
Extracts intent Generates Python Validates against
and entities with pandas/dbt enterprise schema
MCP Code (Ch 13):
# Snowflake MCP server excerpt
class SnowflakeConnector:
async def execute_query(self, query: str, tenant_id: str) -> dict:
# Tenant isolation enforced at connection level
connection = await self.pool.get_connection(
tenant_id=tenant_id,
query_timeout=30
)
result = await connection.execute(query)
return {"rows": result.fetchall(), "schema": result.schema}Observability Code (Ch 17):
# Prometheus metrics for DataForge
CODE_ACCEPTANCE = Counter(
'dataforge_code_acceptance_total',
'User acceptance of generated code',
['user_tier', 'complexity']
)
PIPELINE_SUCCESS = Histogram(
'dataforge_pipeline_success_duration_seconds',
'Time for generated pipeline to first success',
buckets=[60, 300, 900, 3600]
)Focus: Integration (Full Vibe-to-Production) Type: Retail AI startup (series A) Domain: Retail personalization AI Complexity: Medium-High (LLM + recommendations + real-time)
RetailMind helps brick-and-mortar retailers personalize in-store experiences using AI. Their product includes: AI shopping assistants on tablets, personalized promotions, and inventory predictions.
| Chapter | Example Content | Callout Type |
|---|---|---|
| 1 - AI Capabilities | RetailMind evaluates AI for real-time personalization | big-picture: Why retail is a different AI challenge than e-commerce |
| 2 - Synergy Triangle | How RetailMind synchronizes vibe coding, PM, and engineering | key-insight: Full triangle demonstration |
| 3 - Vibe Coding Intro | Weekend prototype: shopping assistant chat interface | practical-example: First 24-hour prototype |
| 4 - Vibe Workflows | Rapid iteration on the recommendation algorithm | code: Prompt engineering for retail persona |
| 5 - Full Lifecycle | Adding analytics and testing to vibe prototype | skill: Testing skill for AI features |
| 6 - Skills | Team-specific skills for deployment and monitoring | skill: Shared team deployment skill |
| 7 - AI PM Strategy | Pivot decision: from B2C to B2B after week 4 | big-picture: When to pivot based on vibes |
| 8 - Eval-First PRDs | Defining "engagement" metrics for in-store AI | document: PRD with multi-metric success criteria |
| 9 - Discovery | User research at retail locations | team-discussion: Store manager interview |
| 10 - Ethics | Privacy concerns with in-store tracking | warning: Customer privacy in physical spaces |
| 11 - GTM | Pilot program with 3 retail chains | document: Pilot program structure |
| 12 - AI Engineering | Production architecture for real-time inference | architecture: Edge + cloud hybrid architecture |
| 13 - Agents | Shopping assistant agent with personality | code: Retail persona prompt template |
| 14 - RAG | Product knowledge base RAG | code: Product embedding pipeline |
| 15 - Model Selection | Balancing cost and quality for in-store devices | document: Edge deployment model comparison |
| 16 - LLMOps | Cost management for retail scale (thousands of stores) | key-insight: AI economics at retail scale |
| 17 - Reliability | Offline mode for stores with poor connectivity | architecture: Graceful degradation design |
| 18 - Integration Playbook | Complete 12-month journey from idea to 100 stores | big-picture: Timeline with key milestones |
| 19 - Case Studies | What RetailMind learned about physical retail AI | key-insight: The importance of store associate buy-in |
| 20 - Team Structures | Cross-functional retail AI team | document: Team composition and roles |
Prototype Code (Ch 3):
# First weekend prototype prompt
"Build a chat interface for a retail shopping assistant.
It should ask about preferences, suggest products, and
show promotions. Use a warm, friendly tone. Add a store
map showing product locations."
Team Discussion (Ch 9):
Store Manager: "I like the idea. But my associates are
already overwhelmed. They can't learn another system."
RetailMind PM: "What if the AI surfaces information
your associates already know, just faster?"
Store Manager: "Maybe. But if it gives wrong info,
customers will blame us, not the AI."
Architecture (Ch 12):
┌─────────────────────────────────────────────────────┐
│ Edge Layer (in-store) │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Shopping │ │ Inventory │ │ Promo │ │
│ │ Assistant │ │ Scanner │ │ Engine │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │
│ │ │ │ │
│ └─────────────────┼─────────────────┘ │
│ │ │
│ ┌───────────▼───────────┐ │
│ │ Edge Inference Server │ │
│ │ (Local model cache) │ │
│ └───────────┬───────────┘ │
└──────────────────────────┼──────────────────────────┘
│ Limited connectivity
▼
┌──────────────────────────────────────────────────────┐
│ Cloud Layer │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ RAG Server │ │ Model │ │ Analytics │ │
│ │ (Products) │ │ Router │ │ Pipeline │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
└──────────────────────────────────────────────────────┘
Focus: Cross-functional (Complete Concept to Deployment) Type: EdTech startup (bootstrapped to seed) Domain: Education/Personalized learning AI Complexity: Medium (LLM + learning analytics + assessments)
EduGen creates AI-powered personalized learning paths for vocational training. Their platform assesses learner knowledge, identifies gaps, and generates customized curriculum.
| Chapter | Example Content | Callout Type |
|---|---|---|
| 1 - AI Capabilities | EduGen's founding story: why AI for vocational training | big-picture: The $30B vocational training market |
| 2 - Synergy Triangle | Founder balancing all three pillars as solo operator | key-insight: Synergy triangle in a 1-person team |
| 3 - Vibe Coding | First prototype: AI-generated lesson plans | practical-example: Weekend hackathon to MVP |
| 4 - Vibe Workflows | Iterating on the assessment engine | code: Adaptive quiz generation prompts |
| 5 - Full Lifecycle | Adding user authentication and progress tracking | skill: EduGen development skill |
| 6 - Skills | Custom skill for curriculum generation workflow | document: SKILL.md for lesson generation |
| 7 - AI PM Strategy | First 100 users: gathering feedback | team-discussion: User interview synthesis |
| 8 - Eval-First PRDs | Defining "learning outcome" metrics | document: PRD with measurable outcomes |
| 9 - Discovery | Research on vocational learner pain points | big-picture: Why workers need AI learning |
| 10 - Ethics | Ensuring equitable access to AI learning | warning: Digital divide in vocational training |
| 11 - GTM | First paid customers: small trade schools | document: Pricing and packaging |
| 12 - AI Engineering | Productionizing the curriculum generator | architecture: Content generation pipeline |
| 13 - Agents | Learner agent that adapts to student pace | code: Learning path agent implementation |
| 14 - RAG | RAG on curriculum standards and industry certifications | code: Certification-aware retrieval |
| 15 - Model Selection | Router balancing quality and cost for content generation | document: Cost per course analysis |
| 16 - LLMOps | Monitoring content quality and learner engagement | code: Quality scoring implementation |
| 17 - Scaling | Growing from 100 to 10,000 learners | key-insight: Scaling challenges in EdTech |
| 18 - Integration Playbook | Complete journey: bootstrapped to seed in 18 months | big-picture: EduGen's funding story |
| 19 - Case Studies | Comparing EduGen to competitors (Coursera, Udacity) | comparison-table: AI approach comparison |
| 20 - Team Structures | Growing the team: first hires | document: Org chart evolution |
| 21 - Future-Proofing | Plans for AI tutors and skill certification | research-frontier: Autonomous learning companions |
Founding Story (Ch 1):
"I was working as an electrician when I realized: the
apprenticeship system hasn't changed in 50 years.
Young workers learn on the job, slowly. I thought:
what if AI could accelerate skilled trades training?"
- Maria Chen, Founder of EduGen
Learning Path Agent Code (Ch 13):
# Learner adaptation logic
def adapt_learning_path(learner_id: str, progress: dict) -> LearningPath:
learner = get_learner(learner_id)
if progress.completion_rate < 0.6:
# Learner struggling - add prerequisites
return LearningPath(
additional_topics=[prerequisite_skills],
pace="slower",
practice_ratio=0.7 # More practice problems
)
elif progress.completion_rate > 0.9 and progress.quiz_scores > 0.85:
# Learner excelling - accelerate
return LearningPath(
skip_review=True,
pace="faster",
advanced_topics=True
)
else:
# On track - standard path
return LearningPath(pace="standard")Pricing Document (Ch 11):
## EduGen Pricing (Year 1)
Individual Learner: $29/month
- Access to all courses
- AI tutoring
- Certification prep
Trade School (up to 50 students): $199/month
- Everything in Individual
- Cohort analytics
- Instructor dashboard
- Custom curriculum import
Enterprise (unlimited): $599/month
- Everything in Trade School
- API access
- Custom AI model fine-tuning
- Dedicated success manager
| Example | Company Type | Domain | AI Focus | Primary Book Focus |
|---|---|---|---|---|
| QuickShip | Startup (2 ppl) | Logistics | Medium | Vibe Coding |
| HealthMetrics | Enterprise SaaS | Healthcare | High | AI PM |
| DataForge | Enterprise | Data/ETL | Very High | AI Engineering |
| RetailMind | Series A Startup | Retail | Medium-High | Integration |
| EduGen | Bootstrapped | EdTech | Medium | Cross-functional |
The following callout types will be used throughout the examples:
| Callout Type | CSS Class | Purpose |
|---|---|---|
| big-picture | callout big-picture |
Strategic context and market opportunities |
| key-insight | callout key-insight |
Core takeaways and lessons learned |
| practical-example | callout practical-example |
Real code and workflows |
| warning | callout warning |
Cautionary notes and pitfalls |
| document | callout info or custom |
PRDs, skills, specs |
| code | example-box |
Implementation examples |
| architecture | example-box |
System diagrams |
| team-discussion | callout |
Conversations between roles |
| skill | callout info |
SKILL.md examples |
| comparison-table | comparison-table |
Feature comparisons |
Chapter 1 (AI Capabilities): All 5 examples introduced with context for why AI fits their domain
Chapter 2 (Synergy Triangle): All 5 examples show how the triangle works in their context
Chapters 3-6 (Vibe Coding): QuickShip primary, RetailMind and EduGen secondary
Chapters 7-11 (AI PM): HealthMetrics primary, EduGen secondary
Chapters 12-17 (AI Engineering): DataForge primary, HealthMetrics secondary
Chapters 18-21 (Integration): All 5 examples in case study format
All examples demonstrate eval-first principles:
- QuickShip: Route accuracy evals
- HealthMetrics: Clinical safety evals
- DataForge: Code acceptance evals
- RetailMind: Engagement metrics evals
- EduGen: Learning outcome evals
All examples address AI economics:
- QuickShip: Route computation cost optimization
- HealthMetrics: Model routing for cost/quality balance
- DataForge: Enterprise-scale inference costs
- RetailMind: Edge computing to reduce cloud costs
- EduGen: Content generation cost per learner