I Produced an Album With AI: Here's What It Taught Me About Working With AI Tools

I just released an album.

I also write software for a living, and I used AI heavily to make the music. Somewhere in the middle of it, I realized the lessons were the same ones I keep re-learning at work with AI coding tools. The medium was different. The mistakes were identical.

Here's the opening track, "Signs," so you know what I'm talking about:

Now the part that actually applies to your day job.

A vague prompt gives you a vague result

Every time I typed something lazy, I got something generic back. "Make it emotional." "Something cinematic." The output was technically fine and completely forgettable.

The takes I kept came from prompts that were specific to the point of being uncomfortable — the exact mood, the instruments, what should not be there, the reference I was chasing.

This is the same thing that happens with an LLM in your editor. "Fix this" gets you a shrug. "This throws on empty input; handle the null case and keep the existing return type" gets you something you can use. The tool isn't the bottleneck. The clarity of what you asked for is.

The tool raises the floor, not the ceiling

AI made it possible for me to produce something that sounds professional far faster than I could have alone. That's real, and I won't pretend otherwise. It raised my floor.

But it did nothing for my ceiling. The difference between a track that's "fine" and one that's mine was entirely in the decisions I made — what to cut, what to push, when to stop. The model doesn't have taste. It has an average. Anything above average was me.

Same in code. AI will get you a working function in seconds. Whether that function belongs in your system, whether the abstraction is right, whether it should exist at all — that's still your ceiling, and the tool can't lift it for you.

The real skill is rejecting

I generated a huge amount of material. Most of it never made the album. The work wasn't producing options — the AI did that endlessly. The work was throwing them away.

That flipped how I think about these tools. When output becomes cheap, curation becomes the job. The value isn't in generating more. It's in the judgment to keep the one good thing and delete the nine that were merely plausible.

If you use AI to write code, you already feel this. It'll happily hand you ten approaches. Knowing which one to keep — and being willing to reject the other nine even though they "work" — is the actual skill now.

The model has no skin in the game

The AI doesn't care if the album is good. It won't lie awake if a track is hollow. It has no reputation attached to the output. I do. My name is on it.

That's the part people miss when they talk about AI "doing the work." It can do the producing. It cannot do the being-responsible. The moment you ship — a song, a feature, a fix — you own it completely, no matter how much of it a tool generated. Accountability doesn't have an API.

So I read every line the model gave me, in music the same way I do in code. Not because it's usually wrong, but because I'm the one answering for it.

The last ten percent is still yours

The AI got me most of the way, fast. But the last stretch — the arrangement choices, the moment a song ends, the silence you leave in — none of that came from a prompt. It came from sitting with the thing and deciding.

That's where the album became an album instead of a folder of generated audio. And it's where your project becomes yours instead of a pile of accepted suggestions. The tool covers the first ninety percent. The last ten is the part only you can do, and it's the part that people actually feel.

AI didn't replace me on this record. It changed where my effort went — away from producing every sound by hand, toward deciding which sounds deserved to stay. That's exactly the shift happening in software: less typing, more judgment.

If you want to hear where all that judgment landed, the full album "No Maps For This" is streaming everywhere: