# Claude Code

> The agentic coding tool that runs in your terminal (and the desktop app). What it is, how to get it running, and the small habits that make it useful from day one

_15 min · intermediate · track: vibe-coding · id: claude-code_

> **Team:** 
>
> [Claude Code](https://code.claude.com/docs/en/quickstart) is what the
> wolfpack reaches for when chat hits its ceiling. You point it at a project
> folder, give it a goal, and it does the work: reads the repo, runs commands,
> edits files, checks itself. This course is what you need to know before you
> let it loose.

You met Claude Code briefly in the [Claude](/course/claude) course as the third mode (Chat, Cowork, Code). This course is the deeper look. From here on, the vibe coding track assumes you have it installed and that you have run a few commands through it.

## What it actually is

**Claude Code is an agent that runs inside a project folder.** Not a chat box, not a side panel, not an IDE plugin. A program you start in a directory, that can then read whatever is in that directory, run commands the same way you would in a terminal, edit files, check its work, and keep going until the task is done.

Two consequences worth holding in your head from the start. First, it is not chat: you are not asking it questions about your code, you are giving it a goal and watching it execute. Second, it is more powerful than the chat agents you have used so far, because it has hands. The "[note on what each mode can touch](/course/claude)" warning from the Claude course applies double here: Claude Code runs commands directly on your computer.

## Get it running

**Install once, then start it from any project folder you want to work on.**

The official quickstart is at [code.claude.com/docs/en/quickstart](https://code.claude.com/docs/en/quickstart) and stays more current than any course can. The short version:

1. Install Claude Code (one command, on macOS, Linux, or WSL on Windows).
2. Open a terminal in a project folder. `cd` to wherever your work lives.
3. Run `claude`. Sign in the first time. From then on, the agent has a session inside that folder.

Claude Code also runs inside the Claude desktop app now, in case the terminal is not your home. Same engine, different surface.

> **Tip:** 
>
> **Try it.** Make a folder (`mkdir ~/projects/first-claude-code-test && cd ~/projects/first-claude-code-test`), run `claude`, and ask: "Create a small webpage that introduces me with a headline, a short bio, and three links. Pick a clean theme." Watch what it does, file by file. Open the result in your browser when it finishes.

## The plan-execute loop

**For anything bigger than a one-line change, start with a plan.** Ask Claude Code to lay out the steps before it writes any code. You read the plan, push back on the assumptions you do not like, then approve. The agent runs.

Plan mode is one keystroke away. Press `Shift+Tab` twice (or type `/plan`) to flip the agent into "draft a plan, do not execute" mode. The model spends its tokens on thinking, not building. When the plan looks right, you tell it to go.

This is the [Start with a plan principle](/course/principles) made concrete. It is also the cheapest move there is. Building the wrong thing costs an hour and a thousand lines of tokens; reading a plan and rejecting it costs a minute.

> **Tip:** 
>
> **Try it.** Pick a small project you have been meaning to start. Run `claude` in an empty folder, press `Shift+Tab` twice, and ask: "Plan an MVP for [your project]. Keep it tight: a single HTML file or a small Express app, no database, deployable in an afternoon." Read the plan. Strike one assumption you disagree with. Then say "go".

## Cost awareness

**Claude Code is the only tool in this bootcamp where cost can climb fast.** Tokens add up when an agent runs autonomously, especially with extended reasoning or long-running tasks. The other tools have flat-rate consumer plans; Claude Code is sold either through Pro/Max subscriptions ($20 or $100-200 a month) or with API billing where every run costs something.

Three habits keep cost honest:

- **Use `/clear` between unrelated tasks.** Stale context that sits in the conversation gets re-read on every turn. `/clear` resets the window.
- **Plan before you build.** Cheaper to throw away a plan than a build.
- **Save `/fast` for when speed actually matters.** It is noticeably more expensive than the default. Default is fine for most things; reach for `/fast` only when the wait is the thing that hurts.

If your team has a higher-tier plan, you can run heavier workloads. If you are on the API-billed path, a starting budget around $50/month covers most personal experimentation.

> **Tip:** 
>
> **Try it.** Run `/clear` once, mid-session, after a longer back-and-forth. Notice how fresh the next exchange feels. Then check how often you actually do this in practice.

## What Claude Code is not for

**It is not a chat agent.** If you want to ask "what is the difference between OAuth and SAML", use [Claude](/course/claude) on the web. The Code surface is for changing files, running commands, and producing real output. Treating it like a chatbot wastes both your time and your tokens.

It also is not a replacement for reading. The agent will edit twenty files in a minute. You are still on the hook for the result. The [you-are-responsible rule](/course/ai-first-mindset) does not pause when the model is fast.

## Hands-on

1. Install Claude Code from [code.claude.com/docs/en/quickstart](https://code.claude.com/docs/en/quickstart). Run `claude` in a real project folder you have lying around (or a fresh empty one). Sign in if it is your first time.

2. Press `Shift+Tab` twice to enter plan mode. Ask the agent to plan a small task you have been putting off:

```
Plan a simple personal homepage as a single HTML file: a headline, a bio paragraph, three links, a clean dark theme.
```

Read the plan. Push back on at least one assumption. Approve.

3. After it finishes, run `/clear` and ask the agent to review what it built and suggest two improvements. Pick one, ask for the change, watch the diff, approve.

## Reflect

- The first time you watch Claude Code edit five files in fifteen seconds, the instinct is either "amazing" or "wait, slow down". Which one is yours, and what does that say about how cautiously you should approve changes today?
- Where in your work this week is something you would have asked a developer to build for you, but you could plan and ship it yourself with Claude Code in a couple of hours?