The Honest Starter Kit for Building With AI

You don't need to be a developer. You need a direction.

Most people who ask me how to get started in building applications assume you need to be a developer or know how to code. It's not necessary anymore, but it obviously does help. All you need is a laptop, a few free accounts, and a lot of trial and error at 11pm. The tools available right now — for free or close to it — are genuinely good enough to build real things. The bottleneck isn't access. It's knowing what the pieces are and what order to pick them up in.

This is the first post in a series on building with AI. Not theory. Not "AI is transforming the industry" fluff. Actual tools, actual decisions, and the specific things I use at Stonefield to automate work that used to eat hours every week. We're starting at the very beginning: what exists, what you need, and why the first question most people obsess over — which AI model is best? — is the wrong place to spend your energy.

The model question matters less than you think right now

ChatGPT, Claude, Gemini — they're all good. Seriously. When you're starting out, the difference between them is not your problem. It's like asking which brand of pen writes best before you've decided what you're writing. Pick one, start using it, and move on.

My personal preference is Claude for longer reasoning tasks and drafting. ChatGPT has the biggest ecosystem of tutorials and plug-ins, so if you're learning by watching YouTube videos, it's probably the path of least resistance. Gemini is solid if you're already in Google Workspace. None of this matters much until you're building something that has specific, real constraints — and by then, you'll know exactly what you need.

The free tiers on all three are enough to start learning and to build simple things. You'll hit limits eventually. When that happens, under $30/month on whichever model you've been using is the right call. Don't pay for anything until you've actually bumped into the wall (or set up your environment using the free models).

Three ways to build — and which ones are worth your time

There are basically three categories of tools for building (with or without AI). They're not interchangeable, and knowing the difference saves you from picking the wrong one and wondering why nothing works.

Workflow tools (n8n, Zapier, Make, etc.) connect apps and automate sequences. Think: a new deal submission hits your inbox → AI summarizes it → summary gets posted to Slack → broker gets an auto-reply. You're not writing code. You're dragging blocks around and pointing things at each other. Zapier is the most beginner-friendly. Make is more flexible. I use n8n because it's open-source, self-hostable, keeps improving fast, and doesn't charge you per task once you're running volume. It has a learning curve, but once it clicks, it clicks.

Workflow tools are the fastest way to get something actually running. If you're new to building anything, start here. I'll do one dedicated n8n post — the basics and a real use case — then point you to the hundreds of great tutorials already out there (honestly, many explain it better than I could). After that, this series moves on to where I spend most of my time: hosted apps.

Internal applications run on your own machine. You open them when you need them, they do their thing, and they shut down when you close the laptop. Useful for personal scripts and experiments, not useful if anyone else needs to access the tool, or if you need it running when you're not at your desk. I don't build much in this category anymore, but it's a fine place to experiment without spending a dollar.

Hosted applications live on the internet. Accessible from any device, any time — including by other people, like your brokers or your ops team. This is where the tools I've built at Stonefield live. More setup upfront, but the ceiling is much higher. This is the main focus of this series.

The kitchen analogy for hosted apps (bear with me)

Building a hosted application means working with three tools you may have heard of but maybe never touched: GitHub, Supabase, and Vercel. Here's what they actually do, explained the way I'd explain it to my 10-year-old.

Think of it like running a kitchen.

Supabase is your pantry. It's where all your ingredients live — your data. Each ingredient has its shelf, and you can always add more. These are the equivalent of client names, deal statuses, rate submissions, conversation history, whatever your app needs to remember. When your app needs information, it goes to Supabase. When something new happens, Supabase stores it. Without a pantry, you're cooking from nothing every single time.

GitHub is your recipe book. It holds your code — the instructions for how everything works. More importantly, it tracks every change you've ever made. Changed a recipe last Tuesday and now it tastes wrong? GitHub lets you flip back to the version that worked. It's also where your code lives before it gets cooked — the source of truth for what your app actually is.

Vercel is your oven. It takes your ingredients from your pantry, grabs the recipes from GitHub, cooks them, and serves the finished dish to whoever's at the table — meaning, it deploys your app and makes it live on the internet. Every time you update your recipe book (push a change to GitHub), Vercel automatically re-cooks and re-serves it. You don't have to do anything manually. The oven just knows.

Together: Vercel runs the app, GitHub stores the code, Supabase stores the data. That's the whole stack — pantry = data, recipe book = code, oven = the thing that serves it. Everything I've built at Stonefield runs on some version of this.

What you actually need to start

Not a lot. Here's the honest list:

A computer. Doesn't need to be new or powerful for basic applications. I did a lot of this previously on my 5 year old laptop before it died.

Two to four hours a week. More is better, especially early on when you're trying to build the mental model. Less than two hours and you'll keep forgetting where you left off. Consistency beats intensity here — two hours every week beats eight hours once a month.

About $30/month. Once you're past experimenting. Free tiers on most of these tools are generous. Vercel's free tier handles real traffic. Supabase's free tier handles real data. GitHub is free for most use cases. The $30 typically goes to your AI model subscription when you start hitting token limits. Budget accordingly.

Patience. Not the inspirational-poster kind. The practical kind; the willingness to hit an error you don't understand, paste it into Claude, read the response, try again, and not spiral into thinking you're doing it wrong. You're not doing it wrong. You're just building something new, and that takes a few iterations. And even if you are doing it wrong, so what? That's how people learn.

The AI handles more of the hard parts than you'd expect. The part it can't handle is you giving up after the first thing that breaks.

Where this series goes from here

Next up: a single, practical post on n8n — what it is, how to set it up, and one real workflow you can clone and run. After that, we're going into hosted apps: how to get GitHub, Vercel, and Supabase talking to each other, and how to build something that actually runs on the internet without touching a terminal if you don't want to.

Every post in this series will be a specific thing you can do, not a concept you have to go figure out elsewhere. If you want to follow along, the rest of the series lives at Stonefield Insights.

The takeaway from this post: stop agonizing over which model to use and start picking a tool. Workflow tools if you want to move fast. Hosted apps if you want to build something others can actually use. The kitchen is already stocked — you just have to start cooking.