Writing Song Prompts That Actually Slap
The five levers of a great style prompt — genre, mood, tempo, instrumentation, vocal style — and how to combine them without confusing the model.
There is a huge gap between pop song, happy and a prompt that gets you release-ready audio on the first try. The difference is not magic vocabulary — it is covering five specific levers the model listens for. Miss one, and the model fills the gap with the most generic choice it knows.
The Five Levers
- Genre — be one level more specific than the obvious:
synthwavebeatselectronic,neo-soulbeatsR&B - Mood — one or two emotional words:
melancholic,triumphant,late-night. Three or more start canceling each other out - Tempo — a number works best:
92 BPM. Words likemid-tempoordrivingwork as a fallback - Instrumentation — name 2-4 instruments you must hear:
warm Rhodes piano, upright bass, brushed drums - Vocal style — gender, register, texture:
breathy female alto,gritty male baritone,gang vocals
Format matters less than coverage: comma-separated tags are the native language of these models, and order is mostly cosmetic. What kills prompts is contradiction — chill aggressive ballad, 170 BPM asks the model to be two songs at once. Pick a lane.
| Weak prompt | Strong prompt |
|---|---|
| sad song | melancholic indie folk, 72 BPM, fingerpicked acoustic guitar, soft male vocals, rainy-day mood |
| cool electronic music | dark melodic techno, 124 BPM, rolling bassline, airy female vocal chops, hypnotic warehouse energy |
| upbeat pop | funky disco pop, 115 BPM, slap bass, bright brass stabs, confident female lead vocals, dancefloor joy |
A trailer-grade style prompt
cinematic epic orchestral, 90 BPM, thunderous taiko drums, staccato string ostinato, heroic brass swells, distant choir, dark triumphant mood, trailer music
All five levers present: genre (cinematic orchestral), tempo (90 BPM), instrumentation (taiko, strings, brass, choir), mood (dark triumphant), and an implicit vocal choice (choir, no lead vocal).
Common Traps
- Genre soup — stacking 4+ genres yields mush. Two genres max, and make one dominant
- Essay prompts — long prose dilutes the signal. Tags, not paragraphs
- Artist names — models generally will not imitate named artists; describe the *sound* instead:
90s Seattle grunge, raw baritone vocals - No tempo at all — the model defaults to a safe mid-tempo, which is rarely what your scene needs
One workflow habit separates hobbyists from pros: change one tag at a time. If a generation is close, swap only the mood word, or only the tempo, and re-run — now you know exactly what each tag does in this genre. Keep a swipe file of style strings that worked; after twenty songs you will have a personal preset library that outperforms any tutorial. These are the same muscles you built in Anatomy of a Great Prompt — specificity, one idea per tag, iterate on the weakest element. Next up: the lyrics box, where structure tags do the heavy lifting.
Five levers, one prompt. Go pull them all and hear the difference. Write a song prompt