How to use Claude Code: a first-week guide to setup, plans, permissions and skills

What Claude Code is and where it runs in 2026

Most people who search how to use claude code have used a chatbot and want to know what changes when the AI can touch their files. Anthropic's documentation describes Claude Code as an agentic coding tool that reads your codebase, edits files and runs commands. Its page on how Claude Code works names three phases, gather context, take action and verify results, repeated until the task is done. You can interrupt the loop at any point.

The overview lists where it runs: the terminal, VS Code and JetBrains IDEs, a desktop app and the web at claude.ai/code, all sharing a project's CLAUDE.md files and settings. What you learn in one carries over. Anthropic's terminal guide says you can use Claude Code even if you have never used a terminal before. Coding knowledge is not required to install it, but it decides how much of its work you can judge.

Installing Claude Code, signing in and what it costs

Anthropic's advanced setup page lists the current requirements: macOS 13.0 or later, Windows 10 1809 or later, Ubuntu 20.04 or Debian 10 or later, 4 GB of RAM, an internet connection and one of Anthropic's supported countries. The recommended native installer is one line pasted from the quickstart. Then open a new terminal and type claude --version to see a version number. Native installs update themselves, while Homebrew and WinGet installs must be upgraded by hand.

The first time you type claude in a project folder, it asks you to log in through the browser. The quickstart accepts a Claude Pro, Max, Team or Enterprise subscription, a Claude Console account with pre-paid API credits, or a cloud provider such as Amazon Bedrock. Credentials are stored after that first login. Typing /login inside a session switches accounts later.

Is it free to try? Anthropic's pricing page says the Free plan does not include Claude Code and every paid plan does. On a subscription, terminal work and chats share one pool, with limits on a rolling five-hour window plus weekly limits. Check current prices on that page, since plans change. The /usage command shows what your sessions have consumed.

Terminal, VS Code or the desktop app as a starting point

The terminal is the full version. Anthropic's VS Code page lists the complete set of commands and skills, tab completion and the exclamation-mark shell shortcut as CLI-only. The VS Code extension, which needs VS Code 1.94.0 or later, offers a subset of commands but adds a graphical panel with inline diffs and plans you can edit before accepting. It is the gentler start for anyone who already edits code there.

One catch: the extension bundles a private copy of Claude Code, so it does not put claude on your PATH. To run claude in the integrated terminal you still need the standalone install. If the terminal itself is the obstacle, the overview describes a desktop app that includes Claude Code and requires a paid subscription. The same project files drive every surface.

A first Claude Code session on a small real project

Pick a small project you can afford to break, such as a personal site. Put it under git and make one commit first. Anthropic's best-practices guide says checkpoints only track edits made through Claude's file tools, not shell commands, and are not a replacement for git. Then open a terminal there, type claude and answer the workspace trust dialog the security page describes.

Start with questions. The quickstart suggests asking what the project does, which technologies it uses and where the main entry point is, and Claude reads the files it needs by itself. Then ask for one small change with a visible result, for example a footer with the current year. Read the diff, ask which files changed and let it commit with a descriptive message. Question, change, review, commit: that is the whole first day.

Giving Claude Code a task with context, a plan and a check

The best-practices guide builds most tips on one constraint: the context window fills up fast, and performance degrades as it fills. So keep one task per conversation, run /clear between unrelated tasks and be specific. The guide turns a request to fix the login bug into one that names the symptom, the likely folder and a failing test to write first. The more the request says, the fewer corrections follow.

The second habit is a plan before code. The guide recommends four phases: explore, plan, implement, commit. In plan mode, entered with Shift+Tab or a prompt starting with /plan, Claude reads and proposes but does not edit your source until you approve. If you could describe the diff in one sentence, the guide says, skip the plan.

The third habit is a check the agent can run. Without a test, a build or a screenshot to compare, the guide warns, you become the verification loop. Ask it to show test output instead of claiming success. Press Esc to stop a drifting run, or Esc twice to rewind to an earlier state, and after two failed corrections clear the session and write a better prompt.

Claude Code permission modes and the safety habits behind them

A permission mode decides what Claude may do without asking. Anthropic's permission-modes page lists six of them. Manual mode, whose config value is default, runs only reads without asking, while acceptEdits also lets file edits through. Plan mode reads and proposes. Auto mode hands approvals to a classifier model, and bypassPermissions skips checks, for isolated containers and virtual machines only.

With Claude Code v2.1.283 or later, the same page says, auto mode is the starting mode for interactive terminal and VS Code sessions. Its classifier blocks risky actions such as code piped from curl, force pushes and production deploys by default. The page also warns that auto mode does not guarantee safety. In your first week, press Shift+Tab until the status bar shows manual mode and approve each command yourself.

Anthropic's security page puts the responsibility on you to review code and commands before approving them. Keep secrets out of the chat and block files such as .env with a Read deny rule, since the permissions page notes that a .claudeignore file has no effect. Treat downloaded repositories as untrusted, because prompt injection, text written to steer an AI assistant, can hide in a README. Whether client or employer code may go into a cloud AI tool depends on your contract and local data-protection law, which differs by country.

CLAUDE.md, slash commands, skills, subagents and hooks

CLAUDE.md is a markdown file in your project root that Claude Code reads at the start of every session. Anthropic's memory page suggests putting there what you would otherwise repeat, such as build commands and conventions. Running /init drafts one from your codebase. Keep it under 200 lines with rules concrete enough to verify, because the page calls it context, not enforced configuration.

A skill is a folder with a SKILL.md file: a description of when to use it, then the instructions. Claude loads it when that description fits, or you type a slash and its name. To add skills to Claude Code, save the folder under ~/.claude/skills for all projects or under .claude/skills in one repository. Anthropic's Agent Skills documentation says only each name and description, about 100 tokens per skill, loads at startup. The anthropics/skills repository on GitHub installs as a plugin marketplace, and its README calls these skills demonstrations to test first.

The same documentation says to use skills only from trusted sources and to audit every file of any other, since a skill can direct Claude to run code. Subagents work in their own context window and return only a summary. Hooks are scripts that run at fixed points, such as a formatter after each edit, and unlike CLAUDE.md they always run.

Practising Claude Code this week, alone or with a course

A week is enough for the basics if each day ends with one finished task. On day one, install the tool, trust one folder and ask questions only. On day two, make three small changes, each reviewed and committed. On day three, draft a CLAUDE.md with /init and cut it down, and on day four, run one task in plan mode with a test that must pass. On day five, write one skill of your own after reading the SKILL.md files of two published ones.

For that reading, the VITON13 SCHOOL catalogue of agent skills is free: 71 skills collected from GitHub on 11 shelves, from coding to security and agent workflow, each with its star count, install command and a short guide, updated on 24 October 2026. The star count is a popularity snapshot, not a rating. Practising alone teaches how to use claude code day to day. It is harder to learn alone where an agent earns its place and where it does damage.

That is what a structured course adds. The Artificial Intelligence course of VITON13 SCHOOL has seven modules, from working with AI deliberately and prompting as engineering to AI agents, your own systems, and safety and responsibility. The agents module covers tools, memory, steps, control and stopping conditions. Its outcomes include a personal prompt library with versions and acceptance criteria and a working agent built for one of the learner's real tasks.