What I tell everyone who asks how to get started with Claude Code. I have walked through this on a dozen calls; now it is one link.
Written for someone who has never opened a terminal. Already comfortable there? Skip to how to talk to it.
The tool is the easy part. The change is in what you ask for and what you do while it works.
A chat window gives you text to copy somewhere. Claude Code is an agent: it reads and writes your files, runs commands, installs things and checks its own work. Copying answers between ChatGPT and Claude is the habit to drop first.
Plain English, the way you would brief a sharp new hire. You do not need to know how it will be done. It does not care about typos.
You stop being the one who does the work and become the one who decides what is worth doing and whether the result is good. That takes taste, not a computer science degree.
“The new mindset is you go into Claude Code and you tell it your goals.”Apr 2026, a student starting from zero
Mac first, Linux second, Windows works but has the most setup friction. Any recent laptop will do for one session. Memory starts to matter when you run several at once: 16 GB is tight, 32 GB or more is comfortable.
A subscription, not pay-as-you-go API credits, which cost far more for the same work. Start on Pro. When you keep hitting the usage limit, move to Max ($100, then $200). I went from $20 to the top tier in three days, and it is still the best money I spend.
Where your work is saved, and your undo button. You never need to learn git commands; Claude runs them.
Wispr Flow or Superwhisper. You talk faster than you type, and you describe things in far more detail out loud. Install it in the first week.
“It’s better to just do 20, and then you’ll get to know the cadence of the limits, and you’ll see why you need 100.”Feb 2026, a friend two days into AI
About twenty minutes. Every step after the second one can be handed to Claude itself.
It is just a window where you type instead of click. On a Mac press ⌘ Space, type “Terminal” and hit Enter. On Windows open PowerShell from the Start menu.
Copy the whole line for your computer, paste it, press Enter. The most common mistake is copying only the web address.
curl -fsSL https://claude.ai/install.sh | bash
irm https://claude.ai/install.ps1 | iex
Windows: if “claude” is not recognized afterwards, close the window and open a new one. Still not found? Add %USERPROFILE%\.local\bin to your PATH under Environment Variables, then reopen. Both first-timers I set up on Windows hit this.
Claude works in whatever folder you start it in, so give each project its own.
mkdir first-project cd first-project claude
Choose “Claude account with subscription” (not the API option), sign in in the browser, and say yes when it asks whether you trust the folder.
From here everything is a conversation. Your first prompts should make Claude finish your setup and teach you as it goes:
I have never used a terminal. Teach me the five commands I will actually need, then help me install Wispr Flow.
Help me create a GitHub account and connect it, so you can save my work there.
“Setting up Claude in the first place is usually the hardest part, because once you have the power of it, then you could ask it to do anything.”Oct 2026, installing it live with two first-timers on Windows
Stuck on anything, including how to use Claude Code itself? Ask it. It knows its own commands, can fix its own setup and will explain what it is doing if you ask.
“Claude Code is the best teacher.”
Dictate. Ramble. Give it the background, what you have tried and what good looks like. More context beats clever wording every time.
“All I’m doing is talking. I’m literally just talking to the terminal all day long.”
Paste screenshots straight into the terminal. Record your screen while you explain a problem and give it the link. A picture of the bug beats a paragraph about it.
When it asks you a question you cannot answer, ask for the options with pros and cons, and its recommendation. Most of the time the answer is “go with your recommendation”.
Some things need you: a signup, a payment, a button in another website. When it gets stuck, ask what you need to do by hand, and do just that.
“If it gets stuck, I just tell it: what manual steps do I need to do?”
It shows its working as it goes. The summary of what it did and what it needs from you is at the end. Lost? Ask “where are we, and what’s left?”
Not a tutorial. Pick something that annoys you or that you have been putting off, small enough to finish in a sitting. Then go round the whole loop once. Everything after that is repetition.
“Install Claude Code, play around with it. Your projects that you’ve been waiting months for, you’ll probably finish in like 48 hours.”Feb 2026, a friend with a side-app idea
Each one makes the next one possible. Most people stall because they try the advanced ones first.
For anything bigger than a small change, have it write the plan first, read it, push back, then approve. Changing a plan is cheap; changing code is not.
Shift+Tab until it says plan mode, or just say “plan this first, don’t build yet”.
Never accept “it should work now”. Ask it to test what it built: run it, click through it in a browser, write automated tests. Tests are how it catches its own mistakes.
“Test this end to end and show me it working.”
Have it commit to GitHub after each thing that works. When something breaks, you go back. Later: a branch per feature, and you approve the merge.
“Commit this and push it to GitHub.”
CLAUDE.md is a file in your project that it reads at the start of every session. Every time it makes the same mistake twice, add a rule. Keep it short.
Run /init inside the project, then “add a rule to CLAUDE.md so you never do that again”.
Beginners queue everything into one long conversation until it gets confused. Finish a task, then start fresh.
/clear, or open a new terminal tab and run claude again.
Waiting is the real cost. While one session works, start a second on something else in another tab or folder. Then three. This is where the time savings actually come from.
A terminal with split panes helps. Orca is the friendliest I know of.
Explained the same process twice? Ask Claude to write it up as a skill: a playbook it loads whenever that job comes up.
“Make this a skill so you do it this way every time.”
It is only as good as what it can see. Give it your call transcripts, exports and documents. Grant access read-only, through the service’s own permission screen, never with a password.
The permission prompts are good while you learn what it does. Once you understand them, let it work on its own in your project folders, and keep the irreversible things (money, live systems, sending to people) behind your approval.
“Start 10, 15, 20 different things at once, so that all you’re doing is ideating. Don’t do the work.”Jul 2026, a business owner who had installed it but never used it
Copying code from a chat into files by hand is the slowest way to use any of this.
One session at a time means most of your day is waiting. Start the next thing.
Ten unrelated requests stacked in one session confuse it. One task, then a fresh start.
When it gets something wrong, fix it and then add the rule to CLAUDE.md or a skill, so it stays fixed.
Done means you saw it work. Ask for the proof.
Ask “where do I put the API key?” and paste it there yourself, never into the chat, never into the project’s files on GitHub.
Anything that writes to your store, your books or your customers gets staged first (in a sheet, a draft, a test copy) and you approve it.
Pick one setup and learn it properly. A new tool every week is a way to avoid the work.
A bigger plan costs less than one afternoon of your time. Upgrade when the limit gets in the way.
Building is cheap now. Ask for the polish, the edge cases and the nice-to-haves in the same go.
“Just don’t stop at 80%. Go to 110% every time.”Apr 2026, a founder whose developers had just started using it
No. You need to know what you want and to recognise when the result is not it. The people who do best are curious and ask good questions, not the ones with the most experience.
Claude Code also runs in a desktop app, in the browser at claude.ai/code and inside VS Code. They are the same agent. I recommend the terminal because everything on this page, and everything on How I build, starts there, but start wherever gets you going today.
To start, yes. You will hit the limit when you are using it well, which is the signal to upgrade. Do not buy API credits for this; the subscription is far cheaper.
Out of the box it asks before anything risky, and you can stop it at any moment with Esc. Work inside a project folder, save to GitHub so any change can be undone, give accounts read-only access, and keep money and live systems behind your approval. In over a year of daily use it has never done anything malicious to me.
The same way you trust a contractor: you check the result. Ask it to test its work and show you, try it yourself, and for anything that matters have a second AI (Codex works well) review it. I stopped reading code in late 2025.
Codex is the closest alternative and fine if you already pay for ChatGPT. Cursor is an editor with AI built in. Lovable and similar builders are quick to start and hard to grow. I use Claude Code for the work and Codex to review it.
No. I use it for research, spreadsheets, bank-statement analysis, server admin, writing and this website. If the job lives on a computer, it can probably help.
Ask Claude Code first; it is the best teacher. Then follow the people building with it on X, and when one session starts to feel slow, read How I build.
Where this goes after a year: two dozen sessions at once, agents that test and review each other, and the bottleneck each step removed.
This page changes as I learn what trips people up. Get an email when it, or anything on How I build, changes.
Last updated . Drawn from thirteen recorded calls where I helped someone get started (Dec 2025 – Oct 2026) and the checklist I emailed a friend. Quotes are mine, lightly edited. Install commands and prices from Anthropic’s docs on the date above; both change, so if a command fails, the current one is at code.claude.com.