Intro: What Claude Code is — and how to start without losing control
Getting Claude Code running takes ten minutes. Learning to direct it takes longer, and the habits you form in the first session are the ones you keep. This intro covers both: what the tool is, how to install it, and how to run a first task so that you stay the one who decides what is true.
Three ways to work with Claude
The same Claude models sit behind all three. What changes is how much the AI does on its own, and how much of your environment it can reach.
Chat is question and answer. You ask, Claude responds, and you carry the result forward yourself: copy it, check it, paste it somewhere. The AI takes no initiative.
Cowork, in the Claude desktop app, is the middle step. You hand Claude a task with several steps and it works through your files and connected tools to complete it. For many people this is the sweet spot.
Claude Code uses the same kind of agent, but you control the environment it works in. It runs in a folder you choose, reads and writes the files there, runs commands on your machine, and follows instructions you write down once and reuse: a CLAUDE.md file with the ground rules for the folder, skills for recurring workflows, hooks for rules that must always hold, and permission modes that decide how much it may do without asking.
The difference is not that Claude Code is smarter. It is that you can shape the system around the model, and the system is where the value is.
It is not only for developers
Claude Code was built for software development, and that is still where it is strongest. But a folder of contracts, a set of monthly reports or a pile of meeting notes is also a folder. Anthropic has described its own legal, finance and marketing teams using Claude Code for their workflows. If you can describe the work in plain language and the material lives in files, Claude Code can help.
What you do need is the ability to judge the result. That is the real prerequisite, and the rest of this series keeps coming back to it.
Where you can run it
| Surface | What it is | Local files? |
|---|---|---|
| Terminal | The full tool. Everything in this series starts here. | Yes |
| Desktop app (Code tab) | Claude Code without the terminal. Visual diffs, several sessions side by side. | Yes |
| VS Code, JetBrains | Claude Code inside your editor. See Part 2. | Yes |
| Web and mobile (claude.ai/code) | Runs in the cloud on a repository, not on your machine. Good for long tasks you check on later. | No |
Your CLAUDE.md, settings and connected tools work the same across all of them. This intro uses the terminal because it shows most clearly what is happening.
What you need
- A paid Claude plan or a Console account. Pro, Max, Team and Enterprise include Claude Code. The free plan does not. Organisations can also run it through their own Amazon Bedrock, Google Cloud or Microsoft Foundry account.
- A terminal. Terminal on Mac (Cmd+Space, type "Terminal"), PowerShell or Windows Terminal on Windows. If you would rather avoid the terminal entirely, the desktop app includes Claude Code.
- A folder to work in. Claude Code treats the folder you start it in as its workspace.
Install
Run one command. On macOS, Linux or WSL:
curl -fsSL https://claude.ai/install.sh | bash
On Windows PowerShell:
irm https://claude.ai/install.ps1 | iex
Homebrew (brew install --cask claude-code) and WinGet (winget install Anthropic.ClaudeCode) also work. The native installer updates itself in the background; the package managers do not. On Windows, installing Git for Windows is recommended so Claude Code can use Bash.
Open a new terminal window and check:
claude --version
If you see a version number, you are done. If something looks wrong, claude doctor prints a diagnosis without starting a session.
The only terminal commands you need
You do not need to learn the terminal to use Claude Code. You need three commands:
| Command | What it does |
|---|---|
cd folder-name |
Move into a folder (cd .. goes up one level) |
ls (Mac, Linux) or dir (Windows) |
List what is in the folder |
claude |
Start Claude Code in the folder you are in |
Once Claude Code is running, you talk to it in plain language, the same way you would in chat.
First run
cd ~/Documents/contracts
claude
The first time, a browser window opens so you can log in. After that, the start screen shows which model you are running, which account you are on and which folder you are in. Check the folder. It is the boundary of what Claude will work with.
Your first task: on material where you know the answer
Do not start with a demo folder or an invented company. Start with your own material, in an area where you already know what the right answer is. That is the only way to tell whether the result is right.
A good first task is one that reads, sorts and summarises, but does not change anything:
Which of the contracts in this folder renew automatically unless we cancel them? Show a table with the last day to cancel, most urgent first.
Before you send it, press Shift+Tab until the mode indicator shows plan. In plan mode Claude investigates and proposes, but writes nothing.
When the table comes back, open two of the contracts yourself and check the rows. Not because the answer is probably wrong, but because this is the habit that matters. Claude Code checks its own work as part of its loop: it rereads, compares and corrects. That catches syntax errors and inconsistencies. It does not catch the clause it never understood, or the contract that was renegotiated by email last spring. You do.
Six keys
| Key | What it does |
|---|---|
| Esc | Stop Claude mid-task. The work so far is kept, and you can redirect. |
| Esc Esc | With an empty prompt, open the rewind menu to go back to an earlier point. |
| Shift+Tab | Switch permission mode (Manual, accept edits, plan, auto). |
! |
Run a terminal command yourself, for example ! ls. Claude sees the output. |
/clear |
Start a fresh conversation. The old one can still be resumed. |
/help |
List the available commands, including your own skills. |
Typing / on its own opens the menu of all commands. That is how you switch model, among other things.
What undo covers, and what it does not
Rewind (Esc Esc) restores the file edits Claude makes with its own editing tools. It does not restore:
- files changed or deleted by shell commands (
rm,mv,cp) - edits made by other sessions, or by you outside Claude Code
- anything that left your machine: an email sent, a ticket updated through a connected tool
So keep the folder under version control with Git. Claude Code can set that up and make the commits for you. Git gives you a history you can always go back to.
But remember what both of them are for. Undo fixes the mistakes you notice. Review is how you find the ones you would otherwise have missed. Neither replaces the other.
Who approves the actions
In recent versions, Claude Code starts in auto mode by default in the terminal and VS Code. In auto mode a second model, a classifier, reviews Claude's actions so you are not asked about each one. That is convenient, and for routine work it is often the right choice.
But be clear about what it means: a model is approving another model's actions. For your first sessions, and for any material that matters, press Shift+Tab and choose Manual (Claude asks before editing files or running commands) or plan. Part 5 covers how to match the mode to the risk.
Where your data goes
Claude Code runs on your machine, but the model does not. When you send a task, your prompt, the file contents Claude reads and the results of the commands it runs are sent to the model. Files Claude never reads are not sent.
This matters for client material and regulated data. The questions are the same as for any AI tool: what does your AI policy allow, which account are you using, and where is the data processed. Organisations that need control over where processing happens can run Claude Code through their own cloud account (Amazon Bedrock, Google Cloud or Microsoft Foundry) and choose the region there. Web sessions on claude.ai/code run in Anthropic's cloud and work on a repository, not on your local files.
The setting does not make the material safe. You do, by choosing what goes in the folder, which mode the agent runs in, and what you check before anything leaves it.
What happens when you type claude
Claude Code loads the folder's context (its structure, any CLAUDE.md file, your settings and connected tools) and gives the model access to read files, run commands and write changes in a loop. Every step is shown in the terminal. Part 1 explains that loop, and why understanding it changes how you write tasks.
Next in this series: Part 1 — The agentic loop: why Claude Code is not autocomplete