How to write a prompt for AI music
The five elements every music prompt needs, what negative prompts fix, and why the second attempt is usually the good one.
A prompt is a brief. Briefs that work are specific about a few things and silent about the rest.
The five elements
- Genre and era — “late-90s trip hop”, not “electronic”
- Mood — two words, no more
- Tempo and energy — a number if you know it, otherwise “slow and sparse”
- Instrumentation — three named instruments beats a paragraph of adjectives
- The voice — register, texture, delivery, or “instrumental”
Together: late-90s trip hop, melancholy, 85 bpm, upright bass, brushed drums, dusty Rhodes, female vocal, low and breathy.
Thirty seconds of writing. Completely different result from “trip hop”.
Name instruments, not feelings
“Warm” and “lush” mean almost nothing to a model. “Upright bass, brushed drums, Rhodes” produces warmth as a consequence.
Adjectives describe the outcome you want. Instruments describe how to get there.
Use negative prompts
If something keeps appearing that you do not want, say so: “no synths”, “no drums until the chorus”, “dry vocal, minimal reverb”.
Most people never try this, then conclude the model cannot do what they want.
Do not prompt for the whole arrangement
Trying to specify every bar produces worse results than specifying the feel and letting the model arrange. You are writing a brief for a session player, not a score.
The exception is structure — that you should control precisely, with section tags.
The second attempt is the good one
The first take tells you whether your prompt was clear. Read what came back, change one thing, run it again.
Changing five things at once means you learn nothing about which one mattered.
Keep the ones that work
Save your five best prompts. Genre, mood, tempo and vocal descriptions are portable across models and across songs — a prompt that produced something good once is a template, not a one-off.