Claude Code Engineering in Practice

From AI Coding Tool User to Agent System Builder

Session Overview

Claude Code is not just a command-line coding assistant — it is a programmable, extensible, and composable agent framework. In this 2-hour session, participants will learn how to engineer reliable AI agents by designing memory systems, delegating tasks to specialized sub-agents, and building reusable skill packages. Each concept is grounded in a hands-on case study drawn from production scenarios.

Format: 2-hour live session (lecture + live demo + discussion)

Target Audience: Developers, tech leads, and AI engineers with basic programming experience

Module 1: Claude Code Architecture & Memory System (30 min)

1.1 Claude Code Tech Stack Overview (15 min)

• Architecture: CLI → Agent → Tool Chain → Extension Mechanisms

• Core capability map: Memory / Sub-Agents / Skills / Hooks / MCP / Headless

• How Claude Code differs from Cursor, Copilot, and other AI coding tools — it is a full agent framework, not just an autocomplete engine

1.2 CLAUDE.md Memory System (15 min)

• The 3-tier memory hierarchy: Project-level / User-level / Global-level

• Teaching Claude your project conventions, coding style, and team agreements

• Case Study: How implementing CLAUDE.md raised code first-pass accuracy and team style consistency

Module 2: Sub-Agents — Task Decomposition & Collaboration (50 min)

2.1 Core Concepts (15 min)

• Why split "one brain" into multiple "specialized roles"

• Isolated execution, permission boundaries, and context management

• Three canonical sub-agent patterns: Read-Only / Filter / Pipeline

2.2 Case Study: The Bug-Fix Pipeline (35 min)

A complete 3-agent pipeline that takes a bug from discovery to verified fix, demonstrating all three sub-agent patterns in one workflow:

• Locator Agent (Read-Only): Scans the codebase with Glob/Grep/Read to pinpoint the bug. No write permissions — cannot accidentally modify code.

• Fixer Agent (Filter): Receives the locator's diagnosis, applies targeted edits with Read/Write/Edit. No execution permissions — cannot run anything.

• Verifier Agent (Pipeline): Runs tests with Bash/Read to confirm the fix. No write permissions — cannot alter the fix.

Discussion: How permission boundaries create reliability — a pattern transferable to any agent framework (LangChain, CrewAI, AutoGen).

Module 3: Skills — Building Reusable Agent Capabilities (30 min)

3.1 SKILL.md Structure & Trigger Mechanism (10 min)

• The description field is a trigger, not documentation — writing it right is critical

• Automatic discovery and on-demand loading: how Skills differ from Slash Commands

3.2 Case Study: Progressive Disclosure Architecture — Financial Analysis Skill (20 min)

A 3-layer architecture that minimizes token waste while maximizing capability:

• Index Layer (~100 tokens): Claude scans skill names and descriptions

• Content Layer (<5K tokens): SKILL.md instructions loaded only when matched

• Appendix Layer (on demand): Referenced templates/standards loaded only when needed

Live Demo: Installing the financial analysis skill and watching Claude automatically discover and apply it when asked to analyze a company's quarterly report.

Wrap-Up & Q&A (10 min)

• Key takeaway: The gap between demo agents and production agents is engineering architecture, not smarter models

• The four pillars of agent reliability: Memory, Isolation, Composability, Automation

• Open Q&A with participants

What Participants Will Walk Away With

• A mental model for Claude Code as a programmable agent framework

• Practical patterns (Memory, Sub-Agents, Skills) applicable to any agent system

• Understanding of how permission boundaries and context isolation create reliability

• A concrete architecture for building reusable, token-efficient agent capabilities