Claude Code and Large-Context Reasoning

O'Reilly Live Learning Course | 4 Hours | Claude Code + MCP + Agents + Skills

Master Claude Code CLI, MCP servers, agentic workflows, and custom skills. This hands-on course takes you from installation to production-ready AI-assisted development workflows.

Course Overview

The arc is Zero -> Context -> Agents -> Hero. Each segment builds on the last.

What You'll Build

A production-ready Claude development environment built on the CLAUDE.md hierarchy:

Prerequisites

• Basic programming knowledge (TypeScript/JavaScript examples used)
• Claude account (free tier works, Pro recommended for API features)
• VS Code or Cursor installed
• Comfortable with terminal/command line basics

Required accounts and tools:

Repository Structure

claude-code/
├── README.md                           # This file
├── CLAUDE.md                           # Claude Code instructions
├── package.json                        # Node.js dependencies
├── .env.example                        # Environment template
│
├── src/
│   └── utils/
│       ├── client.ts                   # Anthropic client factory
│       └── logger.ts                   # Colorful logging utility
│
├── segment_1_quickstart/               # Zero: install + first CLAUDE.md
│   ├── 01_installation.md              # Setup + smallest useful CLAUDE.md
│   ├── 02_verify_setup.ts              # Verify installation
│   ├── 03_terminal_workflows.ts        # Basic CLI workflows
│   └── CLAUDE.md                       # Segment 1 subdirectory rules
│
├── segment_2_context/                  # Context: CLAUDE.md at every scope
│   ├── 01_claude_md_at_every_scope.md  # The lesson the whole course rests on
│   ├── memory_example/                 # Inert teaching example of file-based memory
│   └── CLAUDE.md                       # A worked example at subdirectory scope
│
├── segment_3_agents/                   # Agents: autonomy with boundaries
│   ├── 01_agents_intro.md              # Agent loop, --allowedTools, kill switch
│   ├── 02_agent_loop.ts                # Plan-Execute-Observe loop
│   ├── 03_agent_boundaries.ts          # Permissions and safety
│   ├── 04_custom_agents.md             # Subagents (Part 2)
│   └── CLAUDE.md                       # Boundary spec (the lesson, live)
│
├── segment_4_hero/                     # Hero: skills, subagents, and MCP
│   ├── 01_skills_intro.md              # Skills with dynamic context injection
│   ├── 02_mcp_architecture.ts          # MCP architecture (advanced reference)
│   ├── 02_production_workflows.ts      # Code review, docs, releases
│   ├── 04_mcp_consumption.md           # Consume the microsoft-learn server
│   ├── MCP_SPEC_CHANGELOG.md           # Spec history for advanced learners
│   ├── memory_server/                  # Optional: build-your-own FastMCP
│   └── CLAUDE.md                       # Segment 4 subdirectory rules
│
├── demos/                              # Live walkthrough punchlists
│   ├── segment_1_zero_punchlist.md
│   ├── segment_2_context_punchlist.md
│   ├── segment_3_agents_punchlist.md
│   ├── segment_4_hero_punchlist.md
│   └── mcp-server-demo-punchlist.md    # Optional: build-your-own MCP
│
├── .mcp.json                           # Project MCP wiring (microsoft-learn, DocumentMCP, github)
├── .claude/
│   ├── settings.json                   # Hooks and project settings
│   ├── rules/                          # Project rules for the GenAI-on-Azure track
│   │   ├── python-genai.md
│   │   ├── azure-deployment.md
│   │   ├── secrets-security.md
│   │   ├── testing.md
│   │   └── project-conventions.md
│   ├── commands/                       # Slash commands
│   │   ├── deploy-genai.md             # /deploy-genai: pre-deploy gate + azd deploy
│   │   └── eval-prompts.md             # /eval-prompts: run the prompt-eval suite
│   ├── agents/                         # Subagents
│   │   ├── code-quality-coach.md
│   │   ├── release-manager.md
│   │   ├── claude-code-tutor.md
│   │   ├── python-mcp-expert.md
│   │   ├── terraform-architect.md
│   │   ├── azure-principal-architect.md
│   │   ├── azure-genai-deployer.md     # Deploys Python GenAI apps to Azure, keyless
│   │   └── genai-eval-runner.md        # Runs eval suites before a deploy ships
│   └── skills/                         # Custom skills (canonical location)
│       ├── mcp-scaffold/               # FastMCP server scaffolding
│       ├── claude-md-audit/            # Audits CLAUDE.md hierarchy for drift
│       ├── review-changes/             # Reviews the working tree before commit
│       ├── azure-bicep-skill/          # Azure Bicep authoring
│       ├── azure-ai-deploy/            # Ships a Python GenAI app to Azure, keyless
│       └── genai-prompt-eval/          # Scores GenAI outputs before ship
│
└── docs/
    ├── SETUP.md                        # Detailed setup guide
    ├── TROUBLESHOOTING.md              # Common issues
    └── MCP_REFERENCE.md                # MCP quick reference

Quick Start

1. Clone the repository

   git clone https://github.com/timothywarner-org/claude-code.git
   cd claude-code
   

2. Install dependencies

   npm install
   

3. Configure environment

   cp .env.example .env
   # Edit .env and add your ANTHROPIC_API_KEY
   

4. Verify setup

   npm run segment1:verify
   

5. Run your first workflow

   npm run segment1:workflows
   

NPM Scripts

# Segment 1: Quick Start
npm run segment1:verify      # Verify Claude API setup
npm run segment1:workflows   # Terminal workflow demos

# Segment 2: Context (CLAUDE.md at every scope)
# No npm scripts - the lesson is markdown and live demos. See:
#   segment_2_context/01_claude_md_at_every_scope.md
#   demos/segment_2_context_punchlist.md

# Segment 3: Agents
npm run segment3:agent-loop   # Agent loop demonstration
npm run segment3:boundaries   # Permission boundaries demo

# Segment 4: Hero (Skills, Subagents, and MCP)
npm run segment4:workflows     # Production workflow demos
npm run segment4:architecture  # MCP architecture (advanced reference)
npm run segment4:memory        # Start optional Python FastMCP memory server

# Utilities
npm run mcp:memory           # Start MCP memory server (same as above)
# Note: demos run via `npx tsx` (no build step). There is no root tsconfig.json,
# so `npm run build` (tsc) is intentionally not part of the course workflow.
npm run lint                 # Run ESLint
npm run format               # Run Prettier

Segment Summaries

Segment 1: Zero - Install, CLI, First CLAUDE.md

Cold open: "You just told Claude about your project. Congratulations, it forgot the moment you closed the terminal."

What you'll learn:

• Installing Claude Code (npm install -g @anthropic-ai/claude-code)
• Essential CLI commands, flags, and model aliases
• Writing the smallest useful CLAUDE.md at the repo root
• Verifying Claude reads it on every turn

Key commands:

claude                        # Start interactive session
claude "review this file"     # One-shot query
claude --model sonnet         # Resolves to the latest Sonnet
claude --allowedTools "Read"  # Restrict tools

Segment 2: Context - The CLAUDE.md Hierarchy

Cold open: "You explained your tech stack on Monday. On Tuesday, Claude was guessing again. The fix isn't just one CLAUDE.md - it's CLAUDE.md at the right scope."

What you'll learn:

• The three-scope cascade: user, project, subdirectory
• The @path/file.md import trick for splitting fat CLAUDE.md files
• Five real CLAUDE.md files shipped in this repo as a worked example
• Context engineering as disciplined CLAUDE.md authoring

No new commands - the lesson lives entirely in the markdown hierarchy. Walk the punchlist at demos/segment_2_context_punchlist.md.

Segment 3: Agents - The Loop, Permissions, Subagents

Cold open: "You can let Claude drive ten tools in a row without asking. That is an agent. The same trick that saves you an hour can torch your branch in thirty seconds."

What you'll learn:

• The agent loop: Plan -> Execute -> Observe -> Adjust
• Two autonomy levers: --allowedTools and a CLAUDE.md boundary spec
• Subagents with isolated context windows (Part 2 of the story)
• The CLAUDE.md kill switch you can edit in real time

Permission levels:

claude                                    # Default (confirms each action)
claude --allowedTools "Read,Glob,Grep"    # Read-only tools pre-approved
claude --dangerously-skip-permissions     # Full autonomy (throwaway branches only)

Example agents shipped:

• code-quality-coach - Mentoring agent that leans on the review-changes skill
• release-manager - DevOps agent for versioning and release prep
• claude-code-tutor - Teaching agent for this course
• python-mcp-expert - Pair for FastMCP server building
• terraform-architect - IaC expert for Azure infrastructure
• azure-principal-architect - Azure Well-Architected design and review
• azure-genai-deployer - Deploys Python GenAI apps to Azure via azd, keyless (DefaultAzureCredential, managed identity)
• genai-eval-runner - Runs eval suites (groundedness, relevance, safety, regression) and gates the deploy on a pass/fail verdict

The GenAI-on-Azure track ships as a set alongside these agents: rules in .claude/rules/, the /deploy-genai and /eval-prompts commands in .claude/commands/, and the azure-ai-deploy and genai-prompt-eval skills in .claude/skills/.

Segment 4: Hero - Skills, Subagents, and MCP

Cold open: "A skill is the answer to 'I just typed this exact 12-step prompt for the fourth time.'"

What you'll learn:

• Skills with dynamic context injection (!`bash`), ${CLAUDE_SKILL_DIR}, and context: fork
• Subagents in .claude/agents/ with frontmatter-scoped tools and models
• MCP consumption: query the microsoft-learn server already wired into .mcp.json at the repo root
• Optional homework: build your own FastMCP server (see segment_4_hero/memory_server/)

Skill anatomy (.claude/skills/<name>/SKILL.md is canonical; .claude/commands/ still works for back-compat):

---
name: review-changes
description: Review my uncommitted changes for bugs and missing tests
allowed-tools: Read, Glob, Grep
---

## Current diff
!`git diff HEAD`

## Your task
Flag bugs, anti-patterns, and missing tests.

The !`git diff HEAD` runs at render time, before Claude sees the prompt.

Key Code Examples

Claude API Client Setup

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();  // Uses ANTHROPIC_API_KEY env var

const response = await client.messages.create({
  model: 'claude-sonnet-5',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Review this code for security issues...' }]
});

Consuming an MCP Server (Segment 4)

The microsoft-learn server is pre-wired in .mcp.json at the repo root (the CLI reads project-scoped MCP servers from .mcp.json, not from .claude/settings.json). Inside a Claude Code session:

> /mcp                        # Browse registered servers
> Use the microsoft-learn server to find the GA version of AKS today.

Claude calls microsoft_docs_search, then optionally microsoft_docs_fetch for depth. No server code to write - just consume.

Building Your Own MCP Server (Optional Homework)

# segment_4_hero/memory_server/server.py (excerpt)
from fastmcp import FastMCP

mcp = FastMCP("memory")

@mcp.tool()
def remember_decision(title: str, decision: str) -> str:
    """Persist a project decision to local JSON."""
    # ... save to ./data/memory.json
    return f"Saved: {title}"

Register it:

claude mcp add memory -- bash segment_4_hero/memory_server/start.sh

Advanced Skill with Frontmatter and Dynamic Context Injection

<!-- .claude/skills/pr-summary/SKILL.md -->
---
name: pr-summary
description: Summarize the current pull request
allowed-tools: Bash(gh *)
---

## Pull request context
- PR diff: !`gh pr diff`
- PR comments: !`gh pr view --comments`
- Changed files: !`gh pr diff --name-only`

## Your task
Summarize the PR. Flag risks. Suggest reviewers.

Three gh commands run at render time. Claude only sees the rendered prompt with real PR data inlined.

Claude vs Copilot: When to Use Each

Course Schedule

Learning Resources

Claude Documentation

• Claude Documentation
• Model Context Protocol
• Claude Code CLI
• Anthropic API Reference

Related O'Reilly Content

• How to Prompt Like a Pro - Tim Warner
• GitHub Copilot for Developers - Tim Warner
• Prompt Engineering for Generative AI - James Phoenix & Mike Taylor

Troubleshooting

See TROUBLESHOOTING.md for common issues.

Quick fixes:

# API key issues
echo $ANTHROPIC_API_KEY  # Verify it's set

# Claude Code not found
npm install -g @anthropic-ai/claude-code

# MCP server won't start
npm run mcp:memory

# Dependency issues
rm -rf node_modules && npm install

# Verify setup
npm run segment1:verify

Instructor

Tim Warner - Microsoft MVP (Azure AI and Cloud/Datacenter Management), Microsoft Certified Trainer (25+ years)

• LinkedIn
• Website
• O'Reilly Author Page

License

MIT License - See LICENSE for details.

Questions? Open an issue or reach out via the course chat during live sessions.