An agent in your codebase
Claude Code is an agentic coding tool that understands your codebase, edits your files, runs commands, and integrates with your dev tools — available in the terminal, VS Code, JetBrains, the desktop app, and on the web. Unlike claude.ai, it has direct access to your files, terminal, and codebase, so it does the work rather than handing you code to paste.
Key concept · What "agent" means, and the three things to keep in mind
An AI agent is software that interacts with its environment to reach a goal — at its core, an LLM operating in a loop, with access to tools, services, and even other agents. Claude Code can read and understand your codebase, edit files across the project, run terminal commands and use the output to decide what's next, and search the web. Three things to keep in mind: the context window is finite working memory (so Claude finds answers strategically rather than loading everything); it asks permission by default before running commands or making changes; and it can make mistakes — staying in the loop catches them early.
Explore → plan → code → commit
The one habit to take from the course: don't jump straight to "write the code." Work in four phases.
Key concept · The four phases
Explore & Plan — use plan mode (Shift+Tab until you see "Plan Mode"), where Claude reads but can't edit; it returns a plan you review and refine. This is the cheapest place to course-correct, before any code is written. Code — approve the plan and let Claude work; smooth the phase by defining a success criterion (what "correct" looks like), adding tools (e.g. Claude in Chrome to test a UI), and giving it a reliable test suite to validate against. Commit — before pushing, run a subagent code reviewer for a fresh, unbiased pass, then have Claude generate a commit message in your style.
CLAUDE.md: persistent project memory
Without a CLAUDE.md, Claude Code starts fresh every session — re-exploring the codebase and making assumptions. CLAUDE.md is a Markdown file at your project root that Claude reads automatically each session; its contents are appended to your prompt. Think of it as an onboarding script for your codebase — stack, commands, and code style.
Key concept · A hierarchy, and how to keep it lean
There are two levels: a project-level CLAUDE.md (committed to version control, shared with the team) and a user-level one (in your config folder, personal, across all projects). Tips: save corrections to memory (ask Claude to record a rule you keep repeating), reference docs with @path/to/file, and start without one — see where you keep course-correcting, then run /init to generate a compact, focused file rather than an over-stuffed one.
Subagents, skills, MCP, and hooks
Four ways to extend Claude Code, each with its own job — combine them rather than forcing everything into one.
| Feature | What it does | Reach for it when… |
|---|---|---|
| Subagents | Delegate a task to a separate, isolated context that returns a summary. | Research, code review, or work that would clutter the main context. |
| Skills | Reusable playbooks (SKILL.md) that load on demand when a task matches. | Task-specific expertise you'd otherwise re-explain. |
| MCP | Connect external tools, data, and services via the open protocol. | Claude needs a system or data source outside the codebase. |
| Hooks | Event-driven scripts that fire on file saves, tool calls, etc. | Automated side effects — lint on save, validate before an action. |
Chapters 16 and 17 go deep on skills and subagents; Claude Code in Action (next) covers configuration and unsupervised runs.
Chapter 14 quiz
Fifteen questions on the agent, the workflow, CLAUDE.md, and customization. Answer first, then expand Show answer.
Part A · Multiple choice
What most distinguishes Claude Code from claude.ai?
- A Direct access to your files, terminal, and codebase — it does the work itself
- B It's cheaper
- C It only chats
- D It can't run commands
Show answer
A. It works as an agent, not a paste-back chatbot.
At its core, an AI agent is:
- A An LLM operating in a loop with access to tools
- B A database
- C A static script
- D A search index
Show answer
A.
The recommended workflow is:
- A Explore → Plan → Code → Commit
- B Code → Test → Ship → Hope
- C Plan → Deploy → Explore
- D There's no workflow
Show answer
A.
In plan mode, Claude:
- A Reads but can't edit — it returns a plan to review before any code is written
- B Edits files immediately
- C Deletes the repo
- D Only runs tests
Show answer
A. The cheapest place to course-correct. (Shift+Tab to enter.)
Before committing, it's recommended to:
- A Run a subagent code reviewer for a fresh, unbiased pass
- B Skip review
- C Delete the tests
- D Push without reading
Show answer
A.
What is CLAUDE.md?
- A A Markdown file at the project root Claude reads automatically each session
- B A billing file
- C A compiled binary
- D A test report
Show answer
A. Persistent project memory, appended to your prompt.
The two levels of CLAUDE.md are:
- A Project-level (shared, committed) and user-level (personal, all projects)
- B Free and paid
- C Local and remote
- D Draft and final
Show answer
A.
Which command generates a CLAUDE.md for you?
- A /init
- B /schedule
- C /compact
- D /agents
Show answer
A — /init.
Which customization feature runs in an isolated context and returns a summary?
- A Subagents
- B Hooks
- C CLAUDE.md
- D MCP
Show answer
A — subagents.
Part B · True or false
Claude Code asks permission by default before running commands or making changes.
Show answer
True. You choose how hands-on or hands-off to be.
Giving Claude a reliable test suite helps it know when the code is correct.
Show answer
True. Make sure the tests are a real source of truth (no false positives).
Hooks are request-driven; skills are event-driven.
Show answer
False. It's the reverse — hooks fire on events; skills activate on requests.
Starting without a CLAUDE.md and adding to it over time keeps it lean.
Show answer
True. You learn where course-corrections are actually needed.
Part C · Short answer
Name the four phases of the core workflow and what each does.
Show answer
Explore (gather context), Plan (a plan to measure success), Code (the back-and-forth), Commit (review & push).
Name the four customization features.
Show answer
Subagents, skills, MCP, and hooks (with CLAUDE.md for always-on context).