← AI

Beginning applied AI: Git checkpoints, then Claude Code

There is a version of “getting into AI” that means maths, papers, and training models — I've written a map of that separately. This page is about the other version, the one almost nobody should skip: actually building things with AI. Not watching demos, not reading hot takes — sitting down and making working software with an AI doing the heavy lifting.

The honest pitch: the barrier has collapsed. You need exactly two things — a way to checkpoint your work so nothing can ever really break (that's Git, and you'll use a friendly free app for it, not the command line), and an AI that writes and manages the code while you direct it (that's Claude Code). Everything on this site — the snooker game, the economy charts, the playable MIDI keyboard, the AI chat — was built exactly this way. This page is the whole setup, end to end.

Why you should bother

For decades, having an idea for software and being able to make it were separated by years of training. That gap is now weeks of casual practice. The intimidating parts of programming — the syntax, the arcane tooling, the inscrutable error messages — are precisely the parts the AI now handles. What's left for you is the part that was always the point: deciding what to build, looking at it, and saying “closer, but make it do this instead.” That's product thinking, and you already do it every time an app annoys you.

Which is why passively sitting this out is the one clearly wrong move. You don't need to become an engineer. You need to get over a small hump of unfamiliar words — repository, commit — after which you can build more or less indefinitely. The hump is genuinely small. Here it is.

Step 1: Git, without the intimidation

Git has a fearsome reputation it does not deserve at the level you need it. Strip away the jargon and it is a save system for a folder. Like save points in a video game: every time your project is in a state you like, you save a checkpoint. If anything later goes wrong — you broke something, the AI broke something, you just liked yesterday's version better — you roll back. Nothing is ever lost.

A timeline of commits as save points, with an arrow rolling back from a broken state to the last good checkpoint. first version added a page new feature works something broke roll back — nothing lost
Each commit is a save point for the whole project. The freedom to roll back is what makes it safe to let an AI loose on your code.

And you don't need the command line for any of it. Use GitHub Desktop, a free app for Mac and Windows that turns Git into buttons:

  1. Create a free account at github.com — the place your checkpoints get backed up online.
  2. Install GitHub Desktop and sign in.
  3. In the app, choose New repository. A “repository” is just a folder that Git watches — your project lives in it like any other folder on your computer.
  4. When you've changed something and like the result, the app shows you exactly what changed. Type a one-line note (“added the about page”) and click Commit. That's a checkpoint.
  5. Click Push now and then to copy your checkpoints up to GitHub, so they're safe even if your laptop isn't.

That is the entire Git education you need to start: repository (the watched folder), commit (a save point), push (back it up online). Git can do a great deal more — branches, merges, collaboration — and you should cheerfully ignore all of it for now. The intimidating details can wait, possibly forever; plenty of working developers use little more than this.

Step 2: Claude Code on top

Claude Code is an AI agent that works directly in your project folder. Unlike a chat window you copy code out of, it reads your files, writes new ones, edits existing ones, and runs things — while telling you what it's doing and asking before anything significant. You talk to it in plain English; it does the engineering.

First, a word about the terminal

Claude Code grew up in the terminal, and the word alone puts people off, so let's take its mystique away. The terminal is just a window where you talk to your computer by typing instead of clicking — the original interface, from before mice existed, still there under every Mac and PC. The hacker-movie aesthetic is a costume; in reality it's a polite, empty window that does absolutely nothing until you type something and press Enter. Opening it cannot harm your computer. Looking at it cannot harm your computer. It's a text box.

On a Mac it's the app called Terminal (press ⌘-space and type “terminal”); on Windows it's PowerShell (search the Start menu). Open it and you'll see something like this — and this little session is the entire amount of terminal this workflow ever needs:

A terminal window showing a short session: pwd prints the current folder, cd moves into the project folder, ls lists the files inside, and typing claude starts Claude Code. ~ $ pwd /Users/you ~ $ cd Documents/GitHub/my-website my-website $ ls about index.html style.css my-website $ claude Welcome to Claude Code! > What would you like to build?
A complete terminal session, start to finish: check where you are, step into the project folder, peek at what's inside, start Claude Code. The $ is the prompt — the computer saying “your turn”. You type the highlighted part after it and press Enter; the lines without a $ are the computer's replies.

The one concept that makes all of it click: a terminal is always standing in some folder, exactly like a Finder or File Explorer window is always showing some folder — it just doesn't draw it as icons. Whatever you type happens in the folder you're standing in. And that's the whole reason navigation matters here: when you start Claude Code, it adopts the folder you're standing in as your project — that's what it reads, that's where it writes. Start it in the right folder and it sees your website and nothing else; start it in the wrong one and it's politely confused. So the single job before typing claude is to be standing in your project folder, and three little commands do it:

That's it. pwd to see where you are, ls to look around, cd to walk, Tab to autocomplete — then claude to start. You cannot get lost (you can always pwd), and none of these commands change anything — they only look and move.

And GitHub Desktop will even spare you the walk: with your repository selected, the Repository menu has an “Open in Terminal” item (“Command Prompt” on Windows) that opens a terminal already standing in your project folder. Click that, type claude, Enter. That's the whole ritual.

If even that feels like one step too many, you can skip the terminal altogether: Claude Code also comes as a normal desktop app for Mac and Windows, where you pick your project folder with an ordinary file dialog. Use whichever feels comfortable — they're the same brain in different clothes.

Setting up

  1. Sign up at claude.ai if you haven't — a regular Claude subscription covers Claude Code.
  2. Install Claude Code by following Anthropic's quickstart — either the desktop app or the terminal version. The quickstart is written for beginners; follow it line by line and it works.
  3. Open it in your repository folder — via the app's folder picker, or by typing claude in a terminal opened from GitHub Desktop as above.
  4. The first run walks you through signing in to your Claude account in the browser. After that it just starts.

If anything in the install snags — a confusing prompt, a message you don't recognise — paste it into regular Claude and ask what to do. Using the AI to bootstrap the AI is not cheating; it's the move.

Your first session

Open Claude Code in your repository folder and type something like:

“Create a simple personal website in this folder: a home page about me, plain HTML and CSS, no frameworks. Make it look clean and work on phones.”

It will create the files and explain what it did. Open the page in your browser and look at it — not at the code, at the thing. Then keep going: “make the heading bigger”, “add a page listing my favourite books”, “now make it look good in dark mode”. When you like where it's landed, switch to GitHub Desktop, skim the changes, and commit.

The loop you'll live in

From here on, building software is one loop, repeated:

A four-step loop: describe what you want, Claude Code builds it, you try it and review, then commit a checkpoint and repeat. 1. Describe what you want 2. Claude builds writes & runs the code 3. Try it judge the product 4. Commit save a checkpoint repeat — each pass is small, safe, and reversible
The applied-AI loop. You own steps 1 and 3 — the product judgement. Claude owns step 2 — the code. Git makes step 4 a one-click safety net.

Notice where you spend your attention: on what to ask for and on whether the result is good. The code in the middle is mostly Claude's problem. You'll absorb plenty about how it works over time — by asking, which Claude is endlessly patient about — but understanding every line was never the entry requirement people feared it was.

What this actually gets you

This site is the proof I can offer from my own desk. It's a few hundred pages of plain HTML — no frameworks, no build step — and essentially all of it was built in that loop: in-browser games with their own physics engines (snooker, a 3D RC car simulator), musical instruments and a drum machine, interactive economic charts, live maps, and AI experiments that talk to models through a small server. I directed; Claude built; GitHub Desktop kept the checkpoints. None of it required me to hand-write the code.

Yours doesn't have to be a website. People run this same loop to build small business tools, prototypes for ideas they want to test, scripts that automate tedious work, apps for their family. The loop doesn't care what you point it at.

Habits that make it go well

The fears, answered

Every beginner carries the same handful of worries, so let's retire them explicitly.

No reason to wait

The whole setup — GitHub account, GitHub Desktop, Claude Code, first committed project — is an afternoon, most of it spent watching progress bars. The skills that compound from there aren't the intimidating ones; they're describing what you want clearly and noticing what would make it better. If you've been waiting for the moment when building software stopped requiring you to become someone else first: it's here, and it arrived quietly. Don't watch it go past.

One last thing: this guide is deliberately minimal, which means it can't cover everyone. If it doesn't quite fit your situation — a different machine or setup, a different starting point, a step that didn't work the way I described, or a part you'd like taken further — email me at [email protected] (or say hello on X or LinkedIn) and I'll gladly expand the page. Questions from beginners are exactly the feedback that makes a guide like this better.