Prompting an AI music model is a genuinely different skill from prompting for images, and most people carry the wrong habits across.
With images you describe a scene. With music, describing a scene gets you nothing useful — the model needs to know what the music is, not what it accompanies.
Describe the music, not the moment
The most common failure looks like this:
Music for a video about a startup founder working late at night in a city apartment
Every word describes a scene. Not one describes a sound. The model has to guess the genre, the tempo, the instruments and the production — and it will guess something bland, because bland is the safest average.
Here is the same brief, written as music:
Slow-building electronic instrumental, 95 BPM. Warm analogue synth pads, muted arpeggio, soft sidechained kick entering after 30 seconds. Restrained and hopeful, minimal percussion, lots of space. Modern cinematic production.
Same intent. Completely different result, because every element is now specified.
The five things a music model wants
1. Genre — be specific. "Electronic" is a category containing thousands of unrelated things. "Deep house", "ambient techno", "80s synthwave", "UK garage" each carry a whole set of production conventions the model already knows. Genre is the single highest-leverage word in a music prompt.
2. Tempo. Give a BPM if you know it, or a clear descriptor. 70 BPM and 140 BPM are entirely different pieces of music regardless of what else you specify. Rough anchors: 60–80 slow and heavy, 90–110 mid-tempo, 120–130 dance, 140+ energetic.
3. Instrumentation. Name the instruments. "Brushed drums, upright bass, muted trumpet" produces something specific. "Jazzy instruments" produces an average of everything jazz-adjacent.
4. Mood. Two or three words, no more. Melancholic. Triumphant. Tense. Warm and nostalgic.
5. Production era or reference sound. This is the one people leave out, and it does more work than any other single element. "Recorded to tape, warm saturation, slight wow and flutter" gets you a fundamentally different mix from "modern, clean, wide stereo, heavily compressed".
A template
[Genre] at [BPM], [instrumentation]. [Mood]. [Production character].
Examples:
Lo-fi study beat
Lo-fi hip hop at 72 BPM. Dusty Rhodes piano, brushed drums, warm upright bass, vinyl crackle throughout. Relaxed and slightly melancholic. Recorded to tape, gentle saturation, narrow stereo.
Cinematic trailer
Orchestral trailer cue at 100 BPM building to double time. Low strings ostinato, taiko drums entering at the halfway point, brass swells, choir on the final section. Tense then triumphant. Wide modern film-score production, heavy low end.
Corporate underscore that isn't awful
Minimal electronic underscore at 105 BPM. Muted plucked synth, soft pad, light shaker, occasional piano note. Calm and forward-moving, never busy. Clean modern production, plenty of headroom for a voiceover.
Retro synthwave
80s synthwave at 118 BPM. Analogue bass arpeggio, gated reverb snare, bright lead synth, shimmering pads. Nostalgic and driving. Heavily reverbed, vintage analogue warmth, wide stereo.
Instrumental versus vocal
If you want an instrumental, use the instrumental setting rather than writing "no vocals" in your prompt. A negative instruction in text is unreliable — models frequently produce vocals anyway. The dedicated option is a different code path, not a suggestion.
If you do want vocals, understand what you are asking for: on most models the prompt is a concept, not lyrics to be sung verbatim. You are describing a song; the model writes the words. If you need specific lyrics, look for a model with a dedicated lyrics input, and expect pronunciation to need several attempts.
Writing for video
Music that has to sit under dialogue or voiceover follows different rules from music you listen to directly.
Leave the midrange alone. Voice lives roughly between 200 Hz and 4 kHz. Ask for instrumentation that stays out of the way: sub bass, pads, high sparkle. Prompt for it explicitly — "sparse midrange, leaves room for voiceover".
Ask for less than you think. The instinct is to prompt for something interesting. Under narration, interesting is distracting. "Minimal, repetitive, no dramatic changes" is usually the right brief.
Say where the energy changes. "Builds at the halfway point" gives you something to cut against. A completely flat two minutes gives your edit nothing to work with.
Generate longer than you need. Trimming is free; extending is another generation.
Getting a usable result
Generate several. Music is more variable than images. Three generations of the same prompt will differ more than three images will. Budget for a few attempts and pick.
Change one thing at a time. If a result is close but the tempo is wrong, change only the BPM. Rewriting the whole prompt gives you a different piece of music, not a fixed one.
Steal structural language from the genre. "Four-on-the-floor", "half-time", "call and response", "drop at 1:00" — these are terms the model has learned from real music writing. They carry far more information than "make it more energetic".
When something is wrong, name the fix in production terms. "Too busy" → "sparser, fewer elements, more space". "Too thin" → "add sub bass, wider stereo, warmer low mids". "Too repetitive" → "introduce a new element every 30 seconds".
What AI music is not good at, yet
Being straight about this saves you a wasted afternoon:
- Precise timing to picture. You cannot reliably get a hit on frame 247. Generate, then edit to picture in your NLE.
- Long-form structure. Around two minutes is the practical window. A five-minute arrangement with genuine development is not what these models do.
- Specific lyrics, reliably. Even with a lyrics input, pronunciation and phrasing take several attempts.
- Sounding like a specific artist. Don't — both because it works badly and because it is a legal problem you do not want. Describe the production characteristics you like instead of naming the person.
Rights, honestly
On TurboMax AI you own what you generate, to the extent it is capable of being owned, and we do not use your prompts or outputs to train models.
The genuine caveat is broader than any one platform: AI output may not attract copyright in some jurisdictions, and it can resemble existing work. For background music on a social video that is a manageable risk. For a national ad campaign, get advice first. Anyone who tells you it is entirely settled is selling something.
Generate music on TurboMax AI with prepaid credits — no subscription, the cost shown before you generate, and failed generations refunded automatically.