Claude Code tutorial: a practical guide to getting started

This Claude Code tutorial takes you from install to a first reviewed commit. Anthropic describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools". You need a paid Claude plan or API credits, one install command and a Git repo.

Product facts are as of October 2026; the product changes weekly. Still choosing a tool? Read how Claude Code compares with Cursor and Codex first: the habits below carry over.

In short

• Claude Code has no free plan: the cheapest route is Claude Pro at $20 a month (as of October 2026) or API credits. Install takes one command, then you run claude inside a repo.
• Work in a loop: explore and plan, implement, make Claude run the tests, then review and commit. Since late September 2026, sessions start in auto mode: know what it blocks.
• On Tokn, as of 14 October 2026, 9 postings outside Anthropic name Claude Code, always next to Cursor. The ads test a habit (scoping, delegating, verifying), not a tool.

Claude Code tutorial: what you need before you start

You need a paid plan or API credits, a supported OS and a Git project. The native installer doesn't need Node.js.

Source: Claude pricing. Usage limits reset on a rolling five-hour window, and paid plans add weekly limits on top; /usage shows where you stand. Claude Code runs Anthropic models only.

Per the setup page, you need macOS 13.0+, Windows 10 1809+, Ubuntu 20.04+, Debian 10+ or Alpine Linux 3.19+, and 4 GB of RAM. On native Windows, install Git for Windows too. New to the terminal? Anthropic has a terminal guide for first-time users, and the desktop app includes Claude Code without the CLI.

How to install Claude Code and sign in

Run the official installer for your OS, then run claude in your project folder: a browser window opens to log in.

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

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

:: Windows CMD
curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

The installer shows no progress while it downloads, so wait. Native installs update themselves in the background. brew install --cask claude-code and winget install Anthropic.ClaudeCode also work but don't auto-update. Many tutorials still show npm install -g with Node 18: the npm package now requires Node.js 22 or later, and the docs say not to use sudo.

claude --version
cd /path/to/your/project
claude

If the browser doesn't open, press c to copy the login URL. One trap: if ANTHROPIC_API_KEY is set in your shell and you approve it, Claude Code bills the API instead. If something breaks, run claude doctor, or /doctor in a session.

Your first session on a real repo

Start in plan mode so nothing changes, ask Claude to explain the project, then make one small change and review it.

claude --permission-mode plan

In plan mode Claude reads and answers, but edits wait until you approve a plan. Try the quickstart questions: "what does this project do?", "where is the main entry point?", "explain the folder structure". You don't paste files in; type @ to point at one.

Then pick something small, like a failing test. Approve the plan, read the diff, and ask Claude to "commit my changes with a descriptive message".

The core loop: Claude Code best practices

Anthropic's best practices recommend four phases (explore, plan, implement, commit) and open with one rule: "Give Claude a check it can run: tests, a build, a screenshot to compare." The reason: "Claude's context window fills up fast, and performance degrades as it fills."

Explore and plan

In plan mode, ask Claude to read the relevant code and propose a detailed plan; Ctrl+G opens it in your editor. On approval you choose "Yes, and use auto mode", "Yes, manually approve edits" or "No, keep planning". For small fixes: "If you could describe the diff in one sentence, skip the plan."

Implement with a check Claude can run

"Claude stops when the work looks done", so define done in the prompt:

Add rate limiting to the /login route: at most 5 attempts per minute per IP.
Write tests for the limit and the reset, run the full test suite,
fix any failures, and show me the test output.

"Have Claude show evidence rather than asserting success", then read the tests themselves. GitLab's agentic development postings, advertised in October, named the goal: "Validation that confirms agent-generated code does what was intended, so a passing suite means a working feature and not just a satisfied proxy."

Review and commit

/diff shows the changes and /code-review reviews them in a fresh subagent. Then ask Claude to "commit with a descriptive message and open a PR". Commit before large tasks too: checkpoints only track Claude's file edits, and "This isn't a replacement for git."

Permission modes: what Claude Code can do without asking

Since late September 2026, interactive terminal and VS Code sessions start in auto mode on every plan: Claude acts without asking while a separate classifier model reviews each action.

According to the permission modes page, the classifier blocks anything that "escalates beyond your request, targets unrecognized infrastructure, or appears driven by hostile content Claude read", such as production deploys, force pushes and committing secrets. Anthropic's own warning: "Auto mode reduces permission prompts but does not guarantee safety. Use it for tasks where you trust the general direction, not as a replacement for review on sensitive operations."

Shift+Tab switches modes. We think Manual is the better default while you learn, because you see each command before it runs. Set it in ~/.claude/settings.json:

{
  "permissions": {
    "defaultMode": "default"
  }
}

Then use /permissions to pre-approve safe repeats such as npm run lint. On data: Free, Pro and Max accounts have a setting that allows training on your data, "including when you use Claude Code from these accounts"; Team, Enterprise and API accounts aren't used for training by default. Don't put an employer's code into a personal account without permission.

How to set up CLAUDE.md

CLAUDE.md is a markdown file Claude reads at the start of every session. Run /init to generate a draft from your codebase, then cut it to what Claude can't guess.

Commit ./CLAUDE.md for the project; ~/.claude/CLAUDE.md holds your personal preferences, and ./CLAUDE.local.md local notes you gitignore. For example:

- `npm run dev` starts the app, `npm test` runs the tests
- Run `npm test` before committing
- Use 2-space indentation
- API handlers live in `src/api/handlers/`

Include commands, non-default style rules and gotchas; leave out what Claude can read from the code. The memory docs say to "target under 200 lines per CLAUDE.md file", and warn that Claude treats it "as context, not enforced configuration": a rule that must always run belongs in a hook. If your repo has an AGENTS.md and no CLAUDE.md, Claude Code reads it.

Slash commands worth learning first

Type / to list them all. Beyond /init, /plan, /clear and /rewind, learn these.

Custom commands are now skills: .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy.

Skills, subagents, hooks and MCP: add them when you need them

Each solves a specific problem. Add one when you hit that problem, not on day one.

• Skills: a SKILL.md file in .claude/skills/<name>/ that Claude loads when relevant, or when you call /skill-name. Per the skills docs, "a skill's body loads only when it's used", so long workflows belong here, not in CLAUDE.md.
• Subagents: helpers that each run "in its own context window", for investigations that would fill yours. Custom ones live in .claude/agents/; /agents no longer opens a menu, so ask Claude or edit that folder.
• Hooks: shell commands that run at fixed points, such as a formatter after every edit or a block on protected files. Claude can write one: "Write a hook that runs eslint after every file edit".
• MCP: connections to outside tools and data, added with claude mcp add (below). The MCP docs warn: "Verify you trust each server before connecting it."

claude mcp add --transport http notion https://mcp.notion.com/mcp

On Tokn, as of 14 October 2026, 26 postings at 20 companies name MCP, about 3 times as many as name Claude Code, mostly for building MCP integrations. More in the tools AI job postings ask for.

In your editor: the VS Code extension also installs in Cursor and other VS Code forks; the JetBrains plugin needs the CLI installed.

Common mistakes with Claude Code

The first five come from Anthropic's docs; the last three are ours.

• Kitchen-sink sessions: unrelated tasks in one conversation. /clear between them.
• Correcting over and over: after two failed corrections, /clear and write a better prompt.
• A bloated CLAUDE.md: important rules get lost.
• Trust without verification: no tests, no evidence, no diff review.
• Unscoped exploration: scope the question, or use a subagent.
• Auto mode on sensitive code, or bypass mode outside a container.
• Employer code in a personal account without permission.
• Outdated tutorials: npm with Node 18, an /agents menu, "asks before every edit".

What employers ask for, and a first project to show them

Few ads name Claude Code, and none name it alone. On Tokn, as of 14 October 2026, 10 of about 800 active AI postings in Europe name it: 9 outside Anthropic, at 4 companies, 6 of them near-identical Celonis consulting and deployment roles, close to a forward deployed engineer. That is an anecdote, not a statistic.

All 9 also name Cursor, and 7 name Codex. AI coding tools of any kind appear in fewer than 3% of engineering ads. The ads describe a way of working: Brevo's AI Native Product Engineer ad, advertised in October, wanted "real delegation of substantial work"; three Celonis roles asked for the "Ability to write clear specifications, evaluate AI-generated outputs". No ad names CLAUDE.md. See also the AI skills employers want most.

A first project you can show

1. Pick a real project, yours or an open-source repo you use.
2. Commit a short CLAUDE.md.
3. Ship one feature through the loop: plan, implement with tests, /code-review, small commits.
4. Add one hook and one skill for a task you repeat; optionally, one MCP server.
5. Add a "How I used Claude Code" section to the README: what you delegated, what you changed after review, how you verified it.

In interviews, talk about scoping and verification, not the tool. On a CV, name tools as a set, as the ads do: "coding agents (Claude Code, Cursor)". Our step-by-step guide to getting a job in AI covers the rest.

Find jobs that ask for these skills

Browse LLM and generative AI jobs, then create a job alert so new roles reach you by email. Not a developer? See AI jobs without a technical background.

FAQ

Is Claude Code free?

No. As of October 2026 the Free Claude plan excludes it. The cheapest routes are Claude Pro at $20 a month ($17 billed annually) or API credits through the Claude Console.

Do I need Node.js to install Claude Code?

No, the native installer doesn't need it. Only the npm package does, and it requires Node.js 22 or later.

Does Claude Code work on Windows?

Yes, on Windows 10 1809 or later, natively (PowerShell, CMD or WinGet) or inside WSL. Install Git for Windows for native use.

Is Claude Code safe to run on my repo?

Since late September 2026, sessions start in auto mode, where a classifier blocks risky actions, but Anthropic says it "does not guarantee safety". On unfamiliar or sensitive code, use Manual or plan mode, commit before large changes, and keep bypass mode for isolated containers.

Do employers ask for Claude Code?

Rarely, and never alone. On Tokn, as of 14 October 2026, 9 postings outside Anthropic name it, 6 of them at Celonis, always alongside Cursor and usually Codex.