How to write an AI music prompt that actually works
Most AI music prompts fail for the same reason: they describe a vibe instead of a recording. "Make a chill sad song" gives the model nothing to commit to, so it averages every chill sad song it has ever heard and hands you mush. A good prompt reads like a producer briefing a session — it names decisions, not feelings. Here is the framework we use.
Think in decisions, not adjectives
An adjective like "emotional" can be honoured a thousand different ways. A decision like "felt-hammer piano, no drums, 60 BPM" can only be honoured one way. The more decisions you make, the less the model improvises — and improvisation is where AI music drifts toward generic.
Every strong prompt is built from the same six load-bearing decisions. Make as many of them explicit as you can:
- Tempo — a number in BPM, not "slow" or "upbeat".
- Instrumentation — the two or three instruments that define the palette, named specifically (Rhodes, not "keys").
- Groove / feel — swung, straight, behind-the-beat, four-on-the-floor.
- Vocal direction — gender, register, delivery, doubling, or "instrumental" if none.
- Mix character — warm and mid-forward, bright and airy, dry, tape-saturated.
- Energy & arc — does it build, stay flat, or breathe?
Be specific about the things models get lazy on
Generators have defaults, and the defaults are boring: a generic pop kit, a vocal centred and loud, a build into a chorus. If you don't override a default, you'll get it. So spend your specificity where the model is laziest — usually the drums and the mix.
Instead of "drums", say what kind: "dusty MPC kit slightly behind the beat" or "tight four-on-the-floor kick with shuffled 16th hats". Instead of "good mix", say "vocals tucked into the mix, low end rounded, treble rolled off above 8 kHz". Specific production language is the actual signal the model reads.
Don't name artists — name the technique
Naming a famous artist is tempting and usually backfires: most AI music tools filter artist impersonation, so the prompt either gets ignored or sanitised, and it makes your prompt non-portable between tools. Worse, it outsources your creative decisions to a guess about what that artist "sounds like".
Describe the production characteristics instead. "Sliding 808 with pitch automation", "chipmunk-pitched soul sample on loop", "brushed kit laid back behind the beat" — these are reproducible, portable, and they're what actually shapes the sound.
Keep it tight — the style field is small
Many tools cap the style or description field at around 200–500 characters. That's a feature, not a limit: it forces you to keep only the decisions that change the output. Cut mood words that don't pull weight, cut redundant synonyms, and lead with the decisions a model is most likely to default on.
A useful test: delete any phrase and ask whether the track would come out measurably different. If not, it was decoration — drop it.
A worked example
Weak: "a chill lo-fi beat to study to, kind of sad but hopeful."
Strong: "Tempo 68 BPM. Rhodes electric piano with chorus, brushed kit slightly behind the beat, upright bass, soft vinyl crackle. Mix: warm, mid-forward, treble rolled off, plate reverb on the Rhodes. Mood: melancholy with a thread of hope. Energy: slow burn, no chorus."
Same intent, completely different result — because the second version made decisions the model can't average away.
2,300+ deterministic prompts. Free to try.
Pick your decisions — mood, groove, instruments, energy — and the engine fills the rest in under a second. No card needed.
Try the engine →Frequently asked
How long should an AI music prompt be?
Aim for one or two tight sentences that name your key decisions — tempo, instruments, groove, vocal, mix, energy. Most tools cap the style field around 200–500 characters, so trim anything that wouldn't change the output.
Should I write the prompt as a sentence or a list?
Either works. A labelled run-on ("Tempo: 72 BPM. Instruments: …. Mix: ….") is easy for both you and the model to parse, and makes it obvious which decisions you've actually made versus left to chance.
Why does the same prompt give different results each time?
Generative models are probabilistic — they sample. The more decisions you pin down explicitly, the narrower the range of outputs. If you want reproducibility, use a deterministic prompt builder so the same inputs always produce the same brief.