Auto-Music was my first AI project. The idea was straightforward: produce lyric-style music videos — a track, its words on screen, a fitting backdrop — and automate the making and publishing of them end to end. I built it with almost no engineering background, and it ran.
Where it started
I began with a clear product idea and very little of the technical skill to build it. Rather than wait until I'd learned enough, I treated that gap as the problem to solve: use AI for the parts I couldn't yet write, and pick up the rest by shipping. The goal was a working system in front of reality quickly, not a tidy one.
The wall: copyright
What ended it was a legal constraint, not a technical one. The videos relied on popular, copyrighted songs — which you can produce, but cannot monetise. The rights holders own the recording, and the ad revenue follows them.
I looked seriously at the legitimate route: standing up a small label and licensing music properly. I took that far enough to understand the economics and the operational commitment it required, and concluded it wasn't a business I wanted to build.
Why I sunset it
With no viable path to monetisation and the interesting engineering already solved, I retired it deliberately. It had done its job. It took me from a product idea I couldn't build to a system that ran itself — and that capability, not the videos, was the real return.
Lessons learned
Two things stuck with me.
The first was that the limit had moved. With AI in the loop, what I could build was no longer bounded by my coding ability or my budget — it was bounded by imagination. That reframing changed how I approached everything after it.
The second was that this was, emphatically, not the right way to build software with AI. It worked, but it worked by accident — fragile, and hard to reason about. The question it left me with is the one the rest of my toolchain exists to answer: how do you do this properly?