This is the piece I wish someone had handed me when I started. You do not need a computer science degree to build real software with AI anymore. You need a handful of tools set up the right way, a few habits, and something worth building. Here is the whole setup, in an afternoon.
Five pieces, one loop.
Every AI coding setup worth having reduces to the same loop: you describe what you want, agents in your terminal build it, GitHub keeps the source of truth, Vercel puts it on the internet, and what you learn from the live thing feeds the next round.
Everything below exists to stand this loop up.
The builder's loop. This is the same setup we use at Madrona to ship every product we run, from travel planning to boating guides.
Six steps, one afternoon.
In order, because each step makes the next one easier. GitHub and Vercel are free; the two agents come with subscriptions you may already have. Each step pairs the idea with the actual keystrokes, so you are never left wondering what to type.
Make a free account at github.com. GitHub is where your code lives: every project is a repository, and every change is saved with a note about what changed. You do not need to learn git deeply; your AI tools will do most of the git work for you. You just need the account and the mental model: this is the source of truth.
Open the Terminal app that ships with your Mac. This is where you will talk to your AI agents. You need comfort, not mastery: how to open it, paste a command, and read what comes back. Ask your first agent to explain anything confusing; teaching you is part of its job.
Claude Code is an AI agent that lives in your terminal: it writes the code, runs it, fixes what breaks, and explains what it did. Install it, sign in, open a session in a project folder, and describe what you want in plain language. This one tool is most of the setup. It comes with a Claude Pro subscription.
We also run Codex, a second terminal agent, from OpenAI. Two agents matter more than it sounds: they have different strengths, you can run them side by side on different tasks, and comparing their answers teaches you judgment faster than either alone. It comes with a ChatGPT Plus subscription.
Vercel puts your project on the internet. Sign up with your GitHub account, import your repository, and from then on every change you push goes live automatically. The distance from "it works on my machine" to "here is the link" becomes zero.
Connect the services you use so your agent has reach: it can check a deployment, read the error logs, or draft the follow-up email, instead of you copying and pasting between tabs.
That is the whole setup. A terminal with two AI agents, GitHub holding the truth, Vercel putting it live, and connectors giving your agent reach. Everything else is habits.
Six prompts that do the heavy lifting.
The tools take an afternoon; prompting well is what compounds. These are not magic words, they are working agreements, and every one came from a mistake I stopped repeating. Copy them, then grow your own.
Two minutes reading a plan beats twenty minutes untangling code that solved the wrong problem.
Before you write any code, tell me your plan: what you'll build, which files you'll touch, and anything you're unsure about. Wait for my OK.
This changed my review process more than anything else. The agent catches its own visual bugs instead of me playing QA.
Before you tell me something is done, screenshot the page at phone size and desktop size, look at both yourself, and fix anything that is visibly wrong. Then show me.
The first idea is rarely the best one. Comparing three real directions builds your taste faster than accepting one.
Give me three genuinely different directions for this, not three versions of the same idea. Put them side by side so I can compare, and tell me which you'd pick and why.
When every push goes live automatically, the approval moment is the whole safety system. Guard it.
Never push to the live site without showing me first. Walk me through what changed and wait for my approval.
You are building your own judgment, not just an app. Why beats what, every time.
Explain what you just did and why, in plain language. What would break if we removed it?
Anything you repeat twice belongs in the instructions file. The agent reads it every session, so the project gets easier as it grows.
Create a CLAUDE.md file for this project. Write down the design rules, the things I care about, and the mistakes to avoid. Keep it updated as we learn.
Real, small, and yours.
Do not build a to-do app from a tutorial. Pick something you will actually use this month: a guide to a place you know, a tool for your own work, a site for someone you like. One of our first was a boating guide to the San Juan Islands. A real place, a real need, small enough to ship in a weekend, and it is still live.
Your day-one prompt can be this simple:
I want to build [the thing]: a website that [does what] for [who]. Set up a new project in this folder, keep it simple, and get a first version on screen that I can look at. Then tell me what you did.
Then look at what comes back, react honestly, and go again. That is the whole method.
The tools are the easy part.
Everything above takes an afternoon. What separates people who ship from people who stall is not the setup. It is these.
The questions I get.
GitHub and Vercel have free tiers that will carry you a long way. The agents are the real cost: Claude Code comes with Claude Pro and Codex comes with ChatGPT Plus, each about $20 a month as I write this. Starting with just Claude Code is fine.
Yes. Everything here runs on Windows and Linux too, and both installers cover all three. The guide reads Mac because that is what I work on, but nothing about the loop is Mac-only.
No. You need to describe what you want clearly and look carefully at what you get back. You will absorb more than you expect by asking why, but the agents supply the syntax. You supply the judgment.
The agents ask before running commands or changing files, and git means every committed change is saved and reversible. The honest risk is not a broken computer. It is shipping something you didn't look at, which is what the habits are for.
Commit what works, clear the conversation, and restate the goal in one sentence. If it is still circling, ask the other agent. A second opinion usually breaks the loop, which is half the reason we run two.
An afternoon for the setup, a weekend for a first version you can put in front of someone. The loop is fast. The judgment takes longer, and that part is the fun.
From setup to studio.
This setup is where we started too. Everything else we have written about, the build engine, the agentic operations pattern, grew out of this loop plus time and shipped products. Start the loop, keep it turning, and write down what it teaches you.