Almost every tool that makes music with AI is aimed at the same target: produce a track that sounds like a real record. It is a reasonable target, it is measurable, and the field has got remarkably good at hitting it. It is also not what someone wants when they are making a song for one particular person, and the distance between those two goals is wider than it looks.
Two different definitions of a good song
A generator succeeds when a stranger hears the result and cannot tell how it was made. The audience is everyone, the judgement is production quality, and the way to win is to be closer to the centre of the genre.
A song written for a person succeeds when exactly one listener recognises themselves in it. The audience is one, the judgement is recognition, and being closer to the centre of anything is a way to lose. The two goals do not merely differ; past a certain point they pull against each other.
What the second one is actually made of
The material of a personal song is not sentiment. It is specifics that would be meaningless to anyone else: the argument you always have about the same film, the thing she says before she hangs up, the route home. Sentiment is what you get when those are missing — and a model asked for a song about love will produce sentiment, flawlessly, forever.
Which means the hard part is not generating anything. It is getting the right three facts out of the person who has them and into the song. That is a question of what you are asked, and by what.
A prompt box cannot ask you anything
An empty text field is the most general interface there is, and generality is the problem. It cannot ask who this is for. It cannot ask what happened between you. It gives no reason to mention the parking lot, and so nobody does — they type "a heartfelt song for my wife's birthday, pop" and get back exactly the song that description deserves, which is a birthday song about birthdays.
People do not withhold the good details because they are private. They withhold them because nothing asked, and because a detail that small does not feel like the kind of thing that belongs in a song. It is the only kind of thing that does.
Why the gap has stayed open
Partly because it looks like a small market. "Songs for a person" sounds like birthdays, and birthdays sound like a novelty. But the same shape covers anniversaries, apologies, eulogies, leaving presents, proposals and the thing you make when someone is ill — every occasion where people already spend money on an object that is supposed to mean something, and mostly end up with a card.
And partly because it is the harder engineering problem in the direction nobody gets credit for. Making the model better does not help. Getting a person to tell you the one detail that matters, in under a minute, on a phone, on a day when they left it too late — that is the work, and it looks like nothing in a demo.
What asking properly looks like
Cantio is built around the questions rather than the prompt. It asks who the song is for, what the occasion is and what happened, then writes the lyrics and shows them to you before anything is recorded, so you can change any line that is not quite true. It returns two different songs from the same story, in your choice of language, and the custom song guide walks through it. If you want the craft of it rather than the steps, writing lyrics about someone is the post on choosing the detail.
The thing that gets sent
Nobody forwards a song because the production was clean. They forward it because it mentioned the parking lot. Whatever else these tools get good at, that will keep being supplied by the person who knew about the parking lot — which is the one part of this that was never going to be automated, and the reason the gap is worth closing from the question end rather than the model end.
