Writing Prompts for Music Generation

A text-to-music model turns a description into audio. The single biggest difference between a usable result and a generic one is the description. Think of the prompt as the brief you'd hand a session musician: what kind of music, how fast, which instruments, what it should feel like and how it should unfold.

A prompt is a style brief, not a title

"Summer Nights" tells the model almost nothing. "Warm 80s synth-pop, 112 BPM, gated drums, bright analogue bass, airy female vocal, nostalgic and upbeat" tells it nearly everything it needs. Build prompts out of these parts:

Genre and era

Name the style precisely and, if it matters, the era. "Electronic" covers everything from ambient to hardcore; "melodic deep house" or "late-90s trip-hop with dusty breakbeats" narrows it to a sound.

Tempo

Give a BPM or a narrow range: "around 90 BPM", "174 BPM drum and bass". Tempo shapes the whole feel and is one of the easiest things for a model to honour when you're explicit. If you're not sure of a song's tempo, the free tap BPM counter gets you a number quickly.

Key and mode

If the result has to fit with other material — a vocal you've already recorded, a sample, another track in a set — state the key: "in A minor", "D major". Even if you don't need a specific key, "minor" or "major" is a strong mood signal.

Instrumentation

List the parts you want to hear, with a word about their character:

• Drums: "live acoustic kit", "punchy 808s with rolling hi-hats", "brushed jazz drums".
• Bass: "sub bass", "fingered electric bass", "acid 303 line".
• Harmony and leads: "Rhodes chords", "detuned supersaw pads", "nylon-string guitar arpeggios".

Three to five named parts is usually enough. A list of twelve instruments tends to produce mush.

Mood and energy

Give the feeling and how it changes: "melancholic but hopeful", "tense, building to a euphoric drop", "laid-back late-night groove". Concrete words beat vague ones; "nice" and "cool" carry no information.

Structure and length

Say how the piece unfolds if it matters: "slow intro, full beat at 30 seconds, breakdown, final chorus". Length matters too — a short clip reads as a sketch or loop; something closer to two or three minutes has room for verses and choruses.

Vocals and lyrics are separate from the prompt

On SoundGen's Generate page the prompt describes the sound. Vocals have their own control under it: paste lyrics to have those exact words sung, leave the lyrics empty to let the model write its own, or choose Instrumental for no vocals. A lyric typed into the style prompt is read as part of the description, not sung.

What to leave out

• Real artist, band and song names. Prompts naming them are rejected, for copyright reasons. Describe the sound instead: not "like [a famous band]" but "jangly 80s indie guitar, reverb-heavy vocals, driving eighth-note bass".
• Contradictions. "Minimal and lush", "aggressive lullaby" — the model has to pick one, and you won't like which.
• Mixing instructions. Specific fader moves and plugin settings aren't something a generation model can follow precisely. Fix those afterwards in the Studio.

Iterate deliberately

Treat each generation as an experiment:

1. Start with a clear but not over-specified prompt.
2. Listen for what's right and what's wrong. Write both down.
3. Change one or two things and generate again: tempo, a specific instrument, the mood word.
4. Keep the versions that are close. A strong intro from one take and a better chorus from another can be combined in the Studio.

Example prompts

Lo-fi study beat: "Lo-fi hip-hop, 82 BPM, dusty boom-bap drums with swing, warm Rhodes chords in F minor, soft vinyl crackle, upright bass, relaxed and reflective. Instrumental."

Cinematic cue: "Cinematic orchestral build, 100 BPM, low string ostinato, taiko hits entering halfway, brass swells, rising tension to a big final hit, in D minor."

Dance track: "Melodic techno, 124 BPM, rolling bassline, crisp off-beat hats, evolving arpeggiated synth in A minor, long breakdown with pads, driving and hypnotic."

Notice the pattern: genre, tempo, key, three to five parts, mood, and a hint of structure.