Your first AI song: a 10-minute walkthrough
From an empty prompt box to a finished track with vocals. What to write, which model to pick, and the two mistakes that ruin a first attempt.
The first song most people generate is bad. Not because the models are bad, but because the prompt is one word long and the model has to guess everything else.
Here is the ten-minute version that works.
Write four things, not one
A prompt that says sad pop song gives the model almost nothing. Four elements fix that:
- Genre and era — ”90s alt-rock”, not “rock”
- Mood — what it should feel like, in two words
- Tempo and energy — “slow, sparse” or “120 bpm, driving”
- The voice — who is singing, and how
Put together: 90s alt-rock, wistful, mid-tempo, male vocal with a slight rasp, clean electric guitar, no synths.
That is thirty seconds of writing and it changes the output completely.
Say what you do not want
Negative instructions work. If you keep getting synths in a song that should be acoustic, say “no synths”. If the vocal keeps arriving over-processed, say “dry vocal, minimal reverb”.
Most people never try this and then blame the model.
Pick the model for the genre
They are not interchangeable:
- Pop, hip hop, anything with a tight modern mix — Suno
- Acoustic, jazz, soul, orchestral — Udio
- Instrumental beds and background — Lyria
If you are unsure, run both. It costs one extra render and you learn more from the comparison than from any guide.
The two mistakes
Judging on the first take. The first result tells you whether your prompt was clear, not whether the idea is good. Read what came back, adjust one thing, run it again.
Prompting for a whole album. One song, one prompt. Batch later, once you know what works.
Then keep going
You have audio. You do not have a release.
Cover art and a video come from the same track without exporting anything — that is the part most tools leave to you, and it is usually the part that takes longest.