How to Write AI Music Prompts That Actually Sound Good
Most people write their first AI song prompt in about four words, get something mediocre, and conclude the technology isn't there yet. The technology is usually fine. The prompt was doing almost no work.
Here's what actually moves the needle, in rough order of impact.
The seven variables that matter
1. Genre — but narrower than you think
"Country" covers Johnny Cash, Shania Twain, and Morgan Wallen. That's not a direction, it's a continent. "90s country with pedal steel and a shuffle beat" is a direction. Add a sub-genre, an era, or a signature instrument and you've immediately eliminated most of the wrong answers.
2. Tempo and energy
Give a BPM number or a clear energy description. "Around 75 BPM" and "slow-building, patient" both work; the model handles either. Tempo affects far more than speed — it determines arrangement density, vocal phrasing, and emotional register.
3. Vocal character
This is the most commonly skipped variable and one of the most powerful. Specify:
- Gender and register ("soft female vocals," "deep male baritone")
- Texture ("breathy," "raspy," "smooth," "strained")
- Delivery ("conversational verses, belted chorus," "half-rapped," "whispered")
4. Instrumentation
Name three or four instruments, not fifteen. "Fingerpicked acoustic guitar, upright bass, brushed drums, and a lonesome fiddle" gives the model a clear ensemble. Listing everything you can think of produces mush.
5. Arrangement arc
Tell it how the song should move: "starts intimate with just piano and voice, builds to full band with strings by the final chorus." Dynamics are what make songs feel alive, and the model will happily follow instructions about them — but only if you give them.
6. Subject matter, specifically
"A song about love" is not a subject. "A song about my sister moving to Tokyo and the last dinner we had before she left" is a subject. Specific human situations produce better lyrics for exactly the same reason they produce better human songwriting: detail is what makes emotion legible.
7. Reference feel (carefully)
Describing a style — "in the style of 70s Laurel Canyon folk-rock" or "with the sparse late-night feel of moody R&B" — is useful and legitimate. Asking for a specific person's actual voice is a different thing, and reputable tools won't do it. Aim at the sound, not the identity.
Before and after
Before: "Happy pop song"
After: "An upbeat 118 BPM pop song about finally moving into my own apartment after years of roommates. Bright synths, funky bassline, four-on-the-floor drums. Warm female vocals, conversational in the verses, big layered harmonies in the chorus. Starts sparse, full production by the second chorus."
Same tool. Completely different outcome.
Before: "Sad rap"
After: "A melancholy melodic rap track at 80 BPM about losing touch with your closest friend from school. Muted guitar loop, deep 808 bass, sparse hi-hats. Male vocals blending singing and rapping, reverb-heavy, resigned rather than angry."
Five common mistakes
Writing a paragraph of adjectives with no nouns. "Ethereal, transcendent, emotional, powerful, moving" tells the model almost nothing. Adjectives modify; nouns anchor. Name instruments, tempos, and situations.
Contradicting yourself. "Aggressive but gentle, fast but relaxing" forces the model to average two directions and land somewhere blurry. Pick a lane.
Over-stuffing. Fifteen instruments and four genres in one prompt produces a crowded, unfocused arrangement. Three or four strong constraints beat twenty weak ones.
Ignoring structure. If you never mention how the song should build, you get a flat one. Mention it.
Generating once and giving up. The single biggest behavioral difference between people who get great results and people who don't: iteration. Generate, listen, change one variable, generate again. Two or three passes usually gets you somewhere worth keeping.
A reusable template
Fill in the blanks:
> A [tempo/energy] [specific sub-genre] song about [specific situation]. [Three or four instruments]. [Vocal gender, texture, and delivery]. [How it starts] building to [how it ends]. [BPM] BPM.
That single sentence structure will outperform 90% of prompts people actually write.
Practice on something concrete
Prompting is a skill that improves fast with reps. Try the template above with your own subject, then run it again with one variable changed — the vocal texture, say, or the arrangement arc — and hear exactly what that variable controls.
You can practice free with our AI song generator. If you want professionally tuned prompts to reverse-engineer, every guide in our artist styles collection ends with a complete, specific prompt built on the principles above.