Can you build an app with a single prompt?
You'd think so, from what's out there. Someone ships a SaaS over a weekend. Someone else builds a game in an afternoon. I've been working with AI for about two years now, and I believe a lot of that is real. I also think it leaves out most of the work.
I just shipped a desktop app that AI wrote most of the code for. It took about three months. It turned into a learning exercise on three levels: the software, the product space, and the price. This is where the time went.
what it is
VizAmp is a music visualizer. The Winamp kind. You play music, it turns the sound into a full-screen light show that moves with every beat. Winamp had this in 1999 and nothing on a modern machine quite replaced it. So I rebuilt it.
the fast part
The first working version came together quickly. One large file. About two hundred presets, sorted into categories by guessing from keywords in their names. It opened, it reacted to music, it looked like something.
It was not something I'd ask anyone to pay for.
the design went backwards
My first design was full Winamp chrome. An LCD readout, VU meters, a scrolling ticker, sliders, a CD jewel case holding the album art. I built all of it. Then I used it for a few days.
It felt like a costume. Nostalgia you had to work around to get to the music.
So I threw it out. A second direction went the same way. What shipped is close to the opposite of where I started: minimal controls that get out of the way, and four skins (Modern, Y2K, Retro and Classic99) if you want the nostalgia back.
AI built each of those designs in hours. Deciding they were wrong took days of using them. That part didn't speed up.
the presets were the real job
A visualizer is only as good as its presets. I started by converting a large MilkDrop preset pack, hundreds of them. A big share rendered as black screens.
The cause took a while to find. The only converter available translates the shader code incorrectly, and there was no clean way around it. I built a validator that renders each preset from a cold start and checks the pixels. It caught most of them. Most wasn't good enough for something people pay for.
So I dropped the whole pack and rebuilt the library from presets already in the right format.
Then I watched them. Every one. I cut the ones that didn't hold up, and sorted the rest into moods by hand: Fractals, Fluid, Cosmos, Vortex and the rest. Over 1,100 presets made it in.
No model was going to tell me a preset looks wrong with the bass turned up. That was me, late at night, with the volume up.
the boring part
Then everything that makes an app a product. Code signing and notarization on the Mac. Installers for Windows and Linux, eleven of them. A license key system. A checkout. Delivery emails. Updates that install themselves.
AI helped with every piece. Each piece still needed someone to decide how it should work, test it and catch what was wrong.
the product space
Somewhere in the middle I had to learn who actually wants a visualizer in 2026, and what they expect from one.
Nobody asks about preset file formats. They ask whether it works with Spotify, with Apple Music, with whatever is playing in a browser tab. So VizAmp listens to the computer's audio instead of asking you to load files. They want the feeling they remember without the clutter they don't. That's why the costume came off.
Understanding the software was one thing. Understanding the people it's for was another.
what's fair
The last lesson was the price, and it was the one I thought about longest.
I launched VizAmp at $15, with a plan to raise it to $29 after a launch window. That's how software has been priced for a long time: estimate the effort, add a margin, run a countdown.
A week in, it didn't sit right.
AI has changed what it costs to build software, and the people buying it know that. Charging as if a team had spent a year on it felt out of step with how it was made. But giving it away felt wrong too. Three months of reviewing, rejecting and rebuilding is real work. The judgment is real. The care is real.
So I asked a simpler question. What's fair?
Where I landed is $1.99, one-time, for good. No launch deal. No countdown. Low because building got cheaper. Paid because time and effort still deserve to be compensated, even when a model did the typing.
I don't think I've fully figured this out. I believe a lot of builders are going to have to work through the same question in the next few years. What does a fair price look like when the cost of code drops toward zero, but the cost of judgment doesn't?
what I took from it
AI gets you most of the way there, and it gets you there fast. I wouldn't have built VizAmp on the side of a full-time job without it.
But the last stretch is judgment, taste and review. That's where most of the three months went. It's also the part that separates a demo from something a customer can rely on.
So when I hear "vibe coded" used to write off anything built with AI, I think it misses the point. The tool changed. The standard didn't. And the price should be honest about both.
Not a single prompt. Not a weekend. Three months of deciding what good looks like, and one long look at what fair looks like.
VizAmp is out now for Mac, Windows and Linux. It's $1.99, one-time, with a free 5-day trial. If you want to see what three months of that looks like, it's at vizamp.co.
If you want to follow along with what I build next, you're in the right place.