Claude Code Workshop

Who's in the Room?

• Raise your hand: used any AI coding tool?
• Keep it up: used Claude (chat or API)?
• Still up? Used Claude Code in the terminal?

This helps us calibrate the session for everyone.

Claude Code: Your AI Teammate in the Terminal

• Not autocomplete — a full agent that reads, writes, and runs code
• Lives in your terminal, works with your existing tools
• Understands your entire project, not just one file
• Configurable: teach it your standards and workflows

Under the Hood

• Claude reads files and understands context
• Proposes changes, asks for permission
• Executes: edits files, runs commands, creates commits
• You stay in control at every step

Why Not Just ChatGPT / Copilot / Cursor?

• Terminal-native — works with any editor, any stack, any OS
• Agentic — multi-step reasoning, not one-shot suggestions
• Project-aware — CLAUDE.md gives persistent, shared context
• Extensible — hooks, commands, agent teams
• No vendor lock-in to any IDE

Today's Mission

• Install and configure Claude Code
• Write a CLAUDE.md for YOUR project
• Set up hooks for automated quality checks
• Create custom slash commands
• See multi-agent orchestration in action

By the end: a fully configured AI dev environment you can use on Monday.

The Journey

• Progressive difficulty — everyone starts together
• "Level Up" challenges for those who want more
• Assistants are here to help — raise your hand anytime

How This Works

• Follow along on your machine — this is hands-on
• Green sticky = I'm good  ·  Red sticky = I need help
• Questions anytime, or post to the shared Q&A doc
• If stuck: grab an assistant first

One Command to Install

npm install -g @anthropic-ai/claude-code

• Requires Node.js >= 18
• macOS: may need sudo for global install
• Windows: use PowerShell as Admin, or WSL2 (recommended)

Connecting to Claude

• Option A: Claude Pro/Max subscription (included with plan)
• Option B: API key — export ANTHROPIC_API_KEY=sk-ant-...
• Option C: Workshop key (on your printed card)

# Just run claude -- first run triggers auth
claude

Hello, Claude

$ claude
> What files are in this project and what do they do?

• Claude reads your project files
• Asks permission before any action
• Permission levels: Allow Once, Allow Session, Always Allow

Quick Tip: One-Shot Mode

# Ask a question without entering interactive mode
claude -p "explain the main function in index.js"

# Pipe input for quick analysis
git diff | claude -p "review this diff"

# Process files
cat error.log | claude -p "what went wrong?"

Great for scripting and CI integration.

Your Turn: Install & First Chat

• Install Claude Code
• Authenticate with your workshop key
• Ask Claude: "What is this project about?"
• Try approving a permission request

Without CLAUDE.md

• Claude doesn't know your coding style
• Doesn't know your build commands
• Doesn't know your architecture decisions
• Every session starts from scratch
• Team members get inconsistent results

With CLAUDE.md

• Persistent instructions loaded every session
• Shared with your team via git
• Claude follows YOUR rules, YOUR style, YOUR workflow
• Think: "onboarding docs for your AI teammate"

5 minutes of writing saves hours of corrections.

Three Levels of CLAUDE.md

• All levels merge — more specific overrides more general
• Project-level is the most important for teams

Writing Effective Instructions

Specific. Actionable. Scoped.

Anatomy of a Great CLAUDE.md

# Project: My App

## Tech Stack
- TypeScript, React 19, Vite, Vitest

## Commands
- Dev: `npm run dev`
- Test: `npm test`
- Lint: `npm run lint`

## Code Style
- Functional components only (no classes)
- Use named exports
- Max file length: 200 lines

## Rules
- NEVER modify package.json without asking
- Always add tests for new functions

CLAUDE.md in the Wild

• Monorepo: "When in packages/api, run tests from that directory"
• Legacy: "This codebase uses jQuery. Do not introduce React."
• Team: "We use conventional commits: feat:, fix:, docs:"
• Security: "Never expose env vars in client-side code"

Real examples from production teams.

Creating a CLAUDE.md

• Open starter project in terminal
• Create CLAUDE.md with project-specific rules
• Ask Claude to add a feature — observe it follows the rules
• Break a rule intentionally — show Claude enforcing it

~4 minutes

Write Your CLAUDE.md

• Open the starter project
• Create CLAUDE.md at the project root
• Fill in: tech stack, commands, code style, rules
• Test: ask Claude to create a file

Compare Notes

• Turn to your neighbor
• Compare CLAUDE.md files
• What did they include that you missed?

2 minutes · Then we move on

Shell Commands That Fire Automatically

Configuration

.claude/settings.json     # project-level (commit to git)
~/.claude/settings.json   # global (personal)

• JSON format under the "hooks" key
• Each hook has a matcher (which tool) and command (what to run)
• PreToolUse hooks return JSON: allow, block, or modify

Anatomy of a Hook

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "command": "npx eslint --fix $CLAUDE_FILE_PATH"
      }
    ],
    "PreToolUse": [
      {
        "matcher": "Bash",
        "command": "/path/to/safety-check.sh"
      }
    ]
  }
}

• matcher: regex against tool name (Write, Edit, Bash...)
• $CLAUDE_FILE_PATH: the file being operated on

PreToolUse: The Safety Net

#!/bin/bash
# safety-check.sh -- block dangerous commands
INPUT=$(cat -)
COMMAND=$(echo "$INPUT" | jq -r '.tool_input.command // empty')

if echo "$COMMAND" | grep -qE 'rm -rf|DROP TABLE|format'; then
  echo '{"decision":"block","reason":"Dangerous command blocked"}'
else
  echo '{"decision":"allow"}'
fi

Powerful for teams. Prevent accidental destructive operations.

Hooks You'll Actually Use

• Auto-format: Run Prettier/ESLint after every file edit
• Security scan: Block commits containing secrets or API keys
• Notification: Desktop alert when a long task completes
• Test runner: Auto-run related tests after file changes
• Logging: Record all Claude actions for audit

Start with auto-format -- highest value, lowest risk.

Hooks in Action

• Add PostToolUse lint hook to settings.json
• Ask Claude to create a file with style issues
• Watch hook auto-fix the style
• Add PreToolUse safety hook
• Ask Claude to run a "dangerous" command — blocked!

~5 minutes

Add a Hook

• Create .claude/settings.json in your project
• Add the PostToolUse ESLint hook
• Test: ask Claude to write a file

Custom Commands for Repeatable Workflows

• Markdown files in .claude/commands/ (project) or ~/.claude/commands/ (global)
• Invoke with /command-name in Claude Code
• Template with $ARGUMENTS placeholder
• Built-in: /init, /compact, /review

Think: saved prompts with superpowers. Team-shared via git.

Example: /review Command

# File: .claude/commands/review.md

Review the following file for:
1. Bugs and logic errors
2. Security vulnerabilities (OWASP top 10)
3. Performance issues
4. Code style violations per our CLAUDE.md

File to review: $ARGUMENTS

Provide findings as a numbered list with severity
(critical/warning/info) and suggested fixes.

Usage: /review src/auth/login.ts

Commands Your Team Will Love

• /test-gen <file> — Generate unit tests for a file
• /explain <file> — Explain code for onboarding
• /migrate <desc> — Generate a database migration
• /api-doc <endpoint> — Generate API documentation
• /security-scan — Security review on recent changes

The best commands encode tribal knowledge.

Creating & Using Slash Commands

• Create .claude/commands/ directory
• Write review.md command file
• Use /review src/index.ts in Claude Code
• Show the structured review output

~4 minutes

Create a Slash Command

• Create .claude/commands/ in your project
• Write at least one command
• Test it with /your-command <argument>

When One Agent Isn't Enough

• Context windows have limits
• Some tasks have independent parallel subtasks
• Research and implementation can happen simultaneously
• Complex refactors touch dozens of files

Orchestration Model

• Lead agent breaks down the problem
• Worker agents execute in parallel
• Results flow back for integration

Coordination Primitives

TaskCreate   -- Define a subtask with description + dependencies
TaskUpdate   -- Mark tasks in_progress / completed
TaskList     -- See all tasks and their status
TaskGet      -- Get full details of a specific task
SendMessage  -- Communicate between agents

Built-in tools. Claude Code manages them when you use the Agent tool.

When to Use Agent Teams

Don't over-engineer. Most tasks don't need multi-agent.

Agent Team Refactoring

"Refactor this Express app: convert 5 route files from callbacks to async/await, update tests, update docs"

• Lead agent analyzes the codebase
• Creates tasks for each file + tests + docs
• Worker agents pick up tasks in parallel
• Lead agent verifies and integrates

~8 minutes

Your Turn to Think

• Think of a task from your real work that would benefit from agent teams
• Share with your neighbor
• We'll hear 2–3 examples

2 minutes

Your Mission

• Open the starter project — DevTool CLI utility
• Use Claude Code to build features, configure tools, and ship
• Three checkpoints — go as far as you can
• Assistants are roaming — ask for help

cp -r project ~/claude-workshop/devtool
cd ~/claude-workshop/devtool
npm install
claude

Checkpoints

Show & Tell

• 2–3 volunteers: 60 seconds each
• What did you build?
• What surprised you?

Volunteers picked during the hands-on phase.

What You Learned Today

• Install Claude Code and authenticate
• CLAUDE.md — persistent, team-shared AI instructions
• Hooks — automated quality checks and safety nets
• Slash commands — repeatable custom workflows
• Agent teams — multi-agent orchestration

What to Do Monday Morning

• 1. Add a CLAUDE.md to your main project at work
• 2. Set up one hook (start with auto-lint)
• 3. Create one slash command your team will use
• 4. Commit your CLAUDE.md — share it with your team

Keep Learning

• Docs: docs.anthropic.com/claude-code
• GitHub: github.com/anthropics/claude-code
• Community: Discord / Forum
• Workshop repo: Take it home and keep building

[QR codes for resources + feedback form go here]