A short prompt asks the model to guess everything. A long prompt buries the only line that mattered. The middle ground β six or seven sentences that name a perspective, a scene, and a feeling β is where the generations stop sounding like background music and start carrying weight.
Anchor the perspective first
Before instruments, before tempo, name who is singing and to whom. "A new parent at 3 a.m. talking to a sleeping baby" gives the lyric generator a posture to fall into; "warm acoustic indie" gives it a texture but no aim. The texture follows the posture once the posture exists.
Then place a scene
A single image β kitchen counter, train window, hallway, hotel parking lot β pulls a generic mood into a specific one. You do not need to explain the scene; just hand one over. The model will fill the corners in ways you would not have thought of, and that is the part that makes the result feel found rather than ordered.
Save the production notes for last
Genre, tempo, and instrumentation belong at the end of the prompt, after the model already knows what the song is about. If you lead with "120 BPM, distorted bass, female vocal," the lyric layer has nowhere to grow from. Lead with the scene and posture, and the production notes sharpen what is already there instead of replacing it.
A starting template
A [WHO] on [WHERE], thinking about [WHAT]. The chorus turns when they realize [TWIST]. Mood is [TWO ADJECTIVES]. Production: [GENRE], [TEMPO], [ONE PRODUCTION DETAIL].
This is not a formula to copy verbatim β it is a checklist for the five decisions that have to be in the prompt before the model can do its job. Skip any one of them and the result will feel underdetermined in a way no amount of regeneration can fix.