Most people have had the thought at some point, usually in the middle of something tedious. Why doesn’t something exist that does this? Surely there’s an easier way. I wish I could make something like that.
And usually that’s where it ends. Building software sounds like something for other people, people who studied it or do it all day. So the idea goes back on the shelf.
I think that assumption deserves a second look. Building software hasn’t become easy, and I’ll get to the parts that still take work. But the distance between “I wish this existed” and a first version you can actually try has become much shorter than most people realize.
What changed
For a long time, the technical knowledge needed to turn an idea into software was enough to discourage many people from even trying. AI has changed what that first step can look like.
Today you can start with a conversation. You describe what bothers you, and an AI helps you think it through: what you actually need, what already exists, what a first version could look like. AI coding agents go a step further than a chatbot. They can work inside a project, write the code, run it and tell you what happened. You try something, see what doesn’t work, and adjust.
You don’t need a big project for this either. A small tool that saves you ten minutes a week is a perfectly good place to start.
Where I’m coming from
I’m not starting from zero, and I don’t want to pretend I am. Since I was 14 or 15 I’ve been taking technology apart: setting up private servers for the MMORPGs I played, spending a lot of time on XDA with bootloaders and custom ROMs, modifying nearly every console I owned. I’ve always enjoyed understanding how things work and figuring out what else might be possible. AI didn’t create that curiosity. It changed what I can do with it.
Before Local Transcriber, I’d never worked with Swift, the language behind most Mac apps. I built it because important things said in meetings kept getting lost, and I wanted transcripts made on my own Mac. Later I built GymBook Intel, a small app that reads years of my gym logs and tries to tell me whether I’m actually getting stronger. Coding agents wrote the code for both.
My background helps me notice when something looks off. I don’t think you need it to start.
How I go about it
I usually start with something I want to solve or understand. I talk it through, often with an AI, and ask a lot of questions. What does this need to do? What can I leave out? Slowly the idea gets more specific.
Before I ask an agent to build anything, I try to write down what it should accomplish and how I’ll know it works. Sometimes that’s just a few lines. For more complex projects, I take more time to work through the details. Then the agent does the technical work, and I look at what comes back. I try it, question the choices it made, point out what’s wrong and decide what happens next.
From a thought to a first version
- A thoughtSurely there's an easier way?
- Talk it throughWhat does it need to do? How will I know it works?
- A first versionAn agent does the technical work.
- Try it, adjust itQuestion it, fix what's wrong, decide what's next.
Steps 03 and 04 repeat until it does what it should.
Plans change along the way. With GymBook Intel I started with fifteen questions I wanted answered. Several overlapped, and only two could be answered at first, so I cut them to six. Some ideas turn out not to be worth building at all. That’s normal, and it’s better to find out early.
Why I work this way
I’ve developed a fairly structured way of working over time, but you don’t need to adopt my entire process to get started. A few habits do most of the work.
Being clear about what I want stops me from quickly building the wrong thing, which AI makes very easy. Breaking an idea into small pieces makes it less overwhelming. Checking the result myself is how I find out what actually works. And keeping control of the direction means I end up with something I want, instead of whatever the AI happened to produce.
Checking matters because things do go wrong. When I tested Local Transcriber on one of my own Dutch recordings, about twelve minutes long, it gave me around 168 words and said the job was complete. Other apps got a largely complete transcript from the same file. Software telling you it’s finished doesn’t mean it worked. You still have to look.
It isn’t always smooth, either. Some problems take far longer than they should, and a few I still haven’t solved.
That thing you keep thinking about
So think about the thing you keep wishing existed. Maybe it’s the report you put together by hand every Friday, or the app you’ve pictured but assumed someone else would have to build.
You don’t need to understand everything before you start. Describe the problem to an AI, ask what a very small first version could look like, and try it. You’ll learn as you go, you’ll hit things that don’t work, and you’ll have to decide what you actually want. That part is still yours.
But the thing you’ve been sitting on might be closer than you think.