Ship a real product in 50 minutes, build your first AI agent, then read production code
By the end of this session, you'll have both conceptual understanding and hands-on experience with the foundations of AI.
Build and deploy a working AI-powered product in 50 minutes using Lovable, Bolt, or Claude Artifacts.
Apply the AI Product Canvas: Problem First, 10x Better, One Core Flow, Ship & Learn.
Give Claude real tools and watch it reason about which to use. 25 minutes of hands-on coding.
Walk through NanoClaw's actual source — 700 lines that power 10,500+ GitHub stars. See the theory in practice.
A carefully crafted progression from concepts to hands-on building.
The world changed. You can build in hours what used to take months. Here's how to think about it.
Quick overview of your options. Pick your weapon.
Build, test, and deploy a real AI product. Teams of 2-3.
Gallery walk + top pitches. Celebrate what you built!
The #1 trend in AI right now: agents that take action, not just answer questions.
Build your first AI agent. Real tools. Real reasoning. Real magic.
You built a simple agent. Now see what production looks like.
The transformation narrative and what comes next.
Everything you need to explore, experiment, and build.
Visual explainer: Chatbot vs Agent, the Agent Loop, and the NanoClaw case study. Perfect for review.
View deep-diveNo programming experience needed. Understand how a production AI agent works through business analogies, pseudocode, and just ~20 key lines of real code. Interactive detective challenges included.
Start learningReady for more? Read real production agent code file-by-file. See how index.ts, ipc-mcp.ts, and db.ts map to the Agent Formula. ~700 lines that power 10,500+ GitHub stars.
Explore the codeBuild your first AI agent with 3 tools: calculator, stock lookup, company info. Watch Claude reason about which to use. Requires an Anthropic API key.
Open in ColabSame agent workshop using Google Gemini instead of Claude. Identical tools and structure — just a different LLM backend. Requires a Google API key.
Open in ColabAI-powered full-stack builder. Describe your app, get working code with backend, database, and deployment.
Start buildingInstant web apps from prompts. Great for rapid prototyping and iteration.
Try BoltCreate interactive tools, visualizations, and mini-apps directly in Claude conversations.
Open ClaudeOfficial documentation for giving Claude tools. The foundation of building AI agents.
Read docsThree interactive challenges to build your intuition.
Build and deploy a real AI-powered product in teams of 2-3:
Problem you're solving, target user, 10x advantage over alternatives, one core flow.
Use Lovable, Bolt, or Claude Artifacts. Prompt, iterate, get instructor feedback walking the room.
Swap with another team. Get real user feedback. Fix the #1 issue.
One-click deploy. Prepare your 90-second pitch: Problem, Solution, Demo, Ask.
Showcase your creations and learn from classmates:
Walk around, try every team's product for ~1 minute each.
Pick your favorites: Best Problem-Solution Fit, Most Delightful, Best UX.
Winners do 90-second live demos. Celebrate all teams!
25 minutes in Google Colab — real tools, real reasoning, real magic:
Give Claude a calculator. Ask: 'What is 7,394 × 8,261?' Watch it DECIDE to use the tool.
Add stock lookup + company info. Ask: 'Compare Apple and Microsoft market caps.' Watch Claude chain tools.
Try your own questions. Build a custom tool that would be useful for YOUR product from the build session.
AI Agent = LLM + Tools + Reasoning Loop + Memory
┌─────────────────────────────────────────────┐
│ Chatbot: "What's our Q4 revenue?" │
│ → "I don't have access to that data." │
│ ❌ Dead end. YOU go find it. │
└─────────────────────────────────────────────┘
┌─────────────────────────────────────────────┐
│ Agent: "What's our Q4 revenue?" │
│ → *calls* query_database("revenue", "Q4") │
│ → *calls* calculator("2.3M - 2.05M") │
│ → "Revenue: $2.3M, up 12% QoQ" │
│ ✅ Agent did the work. You got the insight.│
└─────────────────────────────────────────────┘ You built a simple agent. Now see what a production agent looks like — NanoClaw's actual source:
Where Claude decides what to do. The reasoning loop that powers every action.
How Claude talks to the outside world. The MCP protocol that enables any tool.
How Claude remembers across sessions. SQLite keeps context without the cloud.
700 lines total. You could audit this in an afternoon. That's why NanoClaw won.
┌─────────────────────────────────────────────┐ │ OpenClaw (Operator) │ │ • 400,000 lines of code │ │ • Closed source — can't audit │ │ • Summer Yue's inbox deleted │ │ • "Trust us" security model │ └─────────────────────────────────────────────┘ ┌─────────────────────────────────────────────┐ │ NanoClaw │ │ • ~700 lines of code │ │ • Fully open source │ │ • Audit everything in an afternoon │ │ • Built by 1 dev in 1 weekend │ │ • 10,500+ GitHub stars │ │ • Karpathy endorsed │ └─────────────────────────────────────────────┘ "Simplicity beats features. Security is PMF."
Ship fast, learn faster. A working prototype beats a perfect plan. You built a real product in 50 minutes.
Agents = LLM + Tools + Loop + Memory. This equation powers every AI agent in the world. You built one. You understand it.
The AI decides. You didn't write if/else logic. Claude REASONED about which tools to use. That's the breakthrough.
Production code is readable. NanoClaw's 700 lines beat OpenClaw's 400,000. Simplicity wins. Security is product-market fit.
Consumer → Builder → Architect. You went from using AI to building products to understanding production systems in 150 minutes.
Complete these tasks before our next class to prepare for prompt engineering.
Real APIs, error handling, and memory — take your agents from simulated toys to production-ready tools.