Get started with Claude Code

Chapter 01 / 07 · 3 min

What Claude Code changes compared to the chat

The chat answers. Claude Code acts: it opens your folders, changes your files and runs commands on your computer. That is useful, and it is exactly why we go slowly.

Two ways to use it: the Code tab in the Claude app (with buttons), or the claude command in the terminal. It is the same Claude Code. I show you both because after a few days you will want the terminal.

Chapter 02 / 07 · 6 min

Prepare a test folder, with Git

You start in an empty folder that exists only for this. If it goes sideways, you lose nothing.

1. 2.1Step 2.1: Create a brand new folder, for example "test-claude", in Documents.
2. 2.2Step 2.2: Open a terminal in that folder. On Mac: open the Terminal app, type cd followed by a space, drag the folder into the window, then press Enter. On Windows: in File Explorer, type cmd in the folder's address bar.
3. 2.3Step 2.3: Type git --version. A version number appears? Git is already there. If not, install it from git-scm.com.
4. 2.4Step 2.4: Type git init. That is the first snapshot: the folder is now tracked.

Check: type git status. If you read "On branch", you're good. If you read "not a git repository", git init didn't go through: redo step 4 in the right folder.

Chapter 03 / 07 · 8 min

Install Claude Code and check that it answers

Two paths. The first has buttons, the second is the terminal. You can do both, they don't get in each other's way.

1. 3.1Step 3.1: App path: download the app at claude.com/download, install it and sign in with your paid account.
2. 3.2Step 3.2: In the app, click the "Code" tab, choose "Local", then select your test folder.
3. 3.3Step 3.3: Terminal path: in your terminal, paste the command for your system (just below). It shows nothing while it downloads, that's normal.
4. 3.4Step 3.4: When it's done, close the terminal and open a new one. Type claude --version.

# Mac, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

Check: claude --version must print a version number. For a fuller check, type claude doctor: it lists what is fine and what is stuck, without changing anything.

Chapter 04 / 07 · 10 min

Your first session: ask, read, undo

In your test folder, type claude. The first time, it has you sign in through the browser, then asks whether you trust the folder.

$ claude…Quick safety check: Is this a project you created or one you trust?1No, exitYes, I trust this folder

1. 4.1Step 4.1: Choose "Yes, I trust this folder".
2. 4.2Step 4.2: Write: "Create a file hello.txt with one sentence saying hello."
3. 4.3Step 4.3: It shows what it wants to do and waits for your approval. Read it, then say yes.
4. 4.4Step 4.4: Open hello.txt yourself. It exists and says hello? You just validated the first link.
5. 4.5Step 4.5: Undo: type /rewind, pick your message, then "Restore code". The file disappears.

Check that the undo worked: look at your folder, hello.txt is gone. /rewind keeps restore points for the last 100 messages of the session and only tracks files that Claude edited itself. A file changed only by a command has to be put back with Git.

Chapter 05 / 07 · 8 min

Write CLAUDE.md and look at its memory

Claude doesn't remember yesterday. Two things make it remember: a file you write, and notes it takes by itself.

1. 5.1Step 5.1: In Claude Code, type /init. It reads your folder and proposes a CLAUDE.md. Empty folder: it will stay short, that's normal.
2. 5.2Step 5.2: Open the file it created and read all of it. Fix whatever doesn't sound like you.
3. 5.3Step 5.3: Add two or three rules of your own by hand, for example "Answer in plain words, short sentences" or "Ask me before deleting a file".
4. 5.4Step 5.4: Type /memory. You see which memory files are loaded and whether auto memory is on.

Check: close Claude Code, reopen it with claude, and ask "Which rules did you read at startup?". It should quote yours. If it doesn't, the file isn't at the root of the folder where you ran the command.

Chapter 06 / 07 · 10 min

My setup, and what broke

Here is how I work today. Not for you to copy it all: so you see where things go wrong when you stack. Every piece I added worked on its own first. The stacking did the damage.

• A big global CLAUDE.md. My writing rules, my tools, my routes for work. It is read every session, so every line I add is paid for every time. I learned to write short and put the detail in other files.
• Skills. Stored instructions that Claude triggers depending on your request. I have a library of several hundred. What broke: loading everything at once drowns the choice, Claude no longer knows which one to take. I keep a small active set and the rest on disk.
• An automatic hook. A hook is a small script that runs by itself at a precise moment. I had one that cleaned up my texts every time. What broke: it disturbed my subagents and my tests. I put it on standby and run the pass by hand, once, on the finished deliverable.
• Subagents. A subagent goes digging on its own and only hands you a summary, which saves your context. But each one uses quota, and several in parallel eat it faster.
• The cache. At every message, the whole history is sent again. If it's cached, it's almost free. An hour without a message and it expires. Switching model, plugging in a tool or installing a plugin mid-session breaks it too. So I install between sessions, not in the middle of a job.
• Screenshots. One day I had Claude drive a website with screenshots: more than two hundred and seventy in a single session. My plan's limit fell within a few hours. Since then I have it read the page text instead of looking at the image.

The principle I take from it: add one single piece, check that it does what you thought, and only then move to the next.

1. 6.1Step 6.1: Pick one thing to add this week: a rule in CLAUDE.md, or a skill, or a hook. Not two.
2. 6.2Step 6.2: Use it on a real task, in your test folder first.
3. 6.3Step 6.3: Reread the result as if a colleague had done it. What surprised you?
4. 6.4Step 6.4: Nothing odd after a few days? Keep it. Otherwise, remove it.

Chapter 07 / 07 · 2 min

Add Codex if you need it

Codex is OpenAI's coding agent. A ChatGPT account is enough, even a free one (with tight limits), or an API key. It's optional.