How to Use Claude Code (A Beginner's Guide You Can Practice On)

Most people install Claude Code, type something like "make my app better," get a pile of nonsense back, and quietly decide the whole thing is overhyped.

I did that. First real session, I pointed it at a project I actually cared about and gave it a prompt that would have gotten a blank stare from a human junior dev. Then I got annoyed at the tool.

Pot, meet kettle.

Here is the thing nobody tells you on the download page. Claude Code is not autocomplete and it is not chat. It is an agent. It reads your files, makes a plan, edits code, and runs commands on your actual machine. Using it well is a genuinely different skill from using ChatGPT well, and if you drive it like a chatbot you are going to have a bad time.

So let's fix that.

What This Thing Actually Is

Claude Code is a command line tool. You run it inside a project folder and it can see that project the way a new dev on your team would see it.

Which means it can read your files, search the codebase, write new code, run your tests, and make Git commits. You are not copying answers out of a chat window anymore. You point it at a real folder and it goes to work.

And that changes the stakes. A chatbot gives you bad advice. An agent with file access acts on bad advice.

That one sentence is why the habits below matter more than any clever prompt.

Getting It Installed

One command. And no, you do not need Node.js, whatever that old blog post told you.

curl -fsSL https://claude.ai/install.sh | bash

That is the native installer, for Mac, Linux and WSL. Windows PowerShell people, yours is irm https://claude.ai/install.ps1 | iex.

Then move into whatever project you want to work on and fire it up.

cd ~/projects/my-app
claude

First run kicks you to the browser to log in. You need Claude Pro or Max, or API credits on your Anthropic account. There is no free tier here, which surprises people right at the finish line.

Your First Session (Do Not Skip This Bit)

Do not open Claude Code in a project and immediately ask it to build a feature.

Ask it to explain something instead.

What does this project do, and how is it structured?

Looks like a wasted prompt, right? It is not. Claude Code goes and reads your files to answer that, and now the two of you are working from the same understanding of the codebase. Every instruction after that lands better, because it is not guessing at your architecture.

Then hand it something small and real.

The date on the invoice page shows in US format. Users are in the UK.
Fix it, and check whether the same bug exists anywhere else.

Look at what that prompt actually does. It says what is wrong, who it affects, and where to look next. It does not say "fix the date bug," which sends Claude wandering through your entire project hunting for something it cannot identify.

Specific in, useful out. Vague in... well. You know.

The One Habit That Actually Matters

Here is the whole game, and I mean that.

Claude Code works in a loop. It reads, it plans, it acts, then it checks its own work. Read, plan, act, verify.

When the tool feels like magic, it is because that loop closed cleanly. When it feels like a waste of forty dollars, it is almost always because the loop broke at the plan step, and it went charging off editing files before it understood the problem.

Your job is to protect the plan step.

That is basically the entire skill. That is it.

In practice it means you ask for a plan before you ask for code. There is even a mode for it. Hit Shift and Tab twice and you are in plan mode, where Claude investigates and proposes an approach without touching a single file. You get to say "no, not like that" before any damage is done.

That is the difference between a five minute fix and an afternoon spent reverting commits and questioning your life choices.

Write It a CLAUDE.md and Stop Repeating Yourself

Every session, Claude Code starts from nothing. It does not remember that you use tabs, that your tests run with pnpm test, or that the legacy folder is a cursed graveyard nobody is allowed to touch.

So write it down. Once. Put a file called CLAUDE.md in the root of your project.

# Project

Next.js app, TypeScript, Tailwind.

## Commands
- `pnpm dev` to run locally
- `pnpm test` before any commit

## Rules
- Never edit anything in `src/legacy/`
- Use the existing `Button` component, do not write new ones

Claude Code reads that automatically at the start of every single session.

Ten minutes writing this file saves you from repeating yourself for months. It is the highest leverage thing a new user can do and almost nobody does it in week one.

Know Your Permission Modes Before You Need Them

By default, Claude Code asks before it edits a file or runs a command. You can loosen that. You can tighten it. And knowing which mode you are in is what stands between you and a genuinely bad afternoon.

The one that burns people is the mode that stops asking.

It is legitimately useful when you are watching the output and working in a throwaway branch. It is legitimately dangerous when you wander off to make coffee.

The Mistakes Basically Everybody Makes

Asking for too much at once. "Build me a dashboard" is not a task, it is a project. Break it into pieces you could review in one sitting.

Not using Git. Commit before you start anything real. Claude Code can undo its own edits, but Git is the actual safety net and it costs you nothing.

Letting a session run all day. Claude Code has a limited context window. Once a session has been going for hours it starts forgetting what it decided earlier. Run /compact, or just start fresh.

Accepting code you have not read. If you do not understand what it wrote, you cannot maintain it. Ask it to explain the change. It will, happily.

One Last Thing Before I Let You Go

You now know enough to be dangerous in the good way. Install it, ask it to explain your project, work in small steps, plan before you act.

The rest is practice, and practice is the part a blog post genuinely cannot give you. Reading about a terminal command is not the same as running one and watching what comes back.

Which is the whole reason I built the free course. It drops you into a real terminal in your browser and walks you from your first session through to custom skills and hooks, checking your work as you go. Nothing to install to start.

And if you would rather go figure it out yourself, honestly, good. Do that. Just commit before you let it near anything you care about.