Build Sessions

Reading about AI

retention after a week

Building with AI

retention after a week

• Claude Code deep mastery — beyond basics
• Custom CLAUDE.md files per project
• Writing skills (SKILL.md) for your domains
• MCP servers for your internal tools
• Prompt engineering that actually works

---
name: my-domain-skill
description: Expert knowledge for
  my specific domain
---

# Domain Knowledge

## Key Patterns
- Always use X framework for Y
- Our API returns Z format
- Error codes: 4xx = client, 5xx = server

## Code Conventions
- snake_case for Python
- Docstrings on all public functions
- Tests in tests/ mirror src/ structure

• Triggers: when does the skill activate?
• Progressive disclosure: basics first, details on demand
• Domain rules: your team’s conventions baked in
• Examples: show don’t tell

Model Context Protocol — the standard for connecting AI to external systems.

// MCP server example
const server = new McpServer({
  name: "my-tools",
  version: "1.0.0"
});

server.tool("query-db",
  { sql: z.string() },
  async ({ sql }) => {
    const result = await db.query(sql);
    return { content: result };
  }
);

Project Ideas

• Slack MCP — let Claude read/send messages
• Database MCP — query your data naturally
• GitHub MCP — PR reviews, issue triage
• Calendar MCP — schedule management
• Custom API MCP — wrap your internal APIs

• Ollama — easiest way to run models locally
• llama.cpp — raw C++ inference, maximum control
• vLLM — production-grade serving with PagedAttention

When Local Makes Sense

• Privacy: data never leaves your machine
• Cost: after hardware investment, inference is free
• Latency: no network round-trip
• Learning: understand how models actually work

• VPS — Hetzner ($5–20/mo), DigitalOcean, Linode
• Tailscale — mesh VPN, access your machines from anywhere
• Docker — containerize everything
• Nginx/Caddy — reverse proxy, SSL

Total cost: $20–50/month for a powerful personal setup.

End-to-end AI-assisted projects. Use the tools we’ve studied to ship something real in 2–3 sessions.

• Define the problem clearly
• Design with AI (wireframes, architecture)
• Build with Claude Code + Cursor
• Deploy to Vercel / Fly.io / your VPS
• Share with the group

Build tools that track the AI supply chain. Turn our technical understanding into investment edge.

• HBM demand tracker — monitor SK Hynix, Samsung capacity
• DRAM price monitor — spot prices, contract prices, trends
• Earnings call analyzer — extract key metrics from NVIDIA/TSMC calls
• Capex tracker — hyperscaler spending trends

Patterns That Work

• System prompts: set role, constraints, format
• Few-shot examples: show 2–3 examples of desired output
• Chain-of-thought: “Think step by step”
• Structured output: ask for JSON, tables, specific formats
• Persona: “You are an expert in X”

Anti-Patterns

• Vague instructions — be specific
• No examples — show, don’t tell
• Too many constraints — models juggle poorly
• Ignoring context window — relevant context > more context
• Not iterating — first prompt is rarely best

Ground Rules

• Bring your laptop. This is a lab, not a lecture.
• Pick a track or go off-script. Both are fine.
• Help each other. Pair programming encouraged.
• Ship > perfect. Deploy something, even if it’s rough.
• Document as you go. Future you will thank you.