Read the prompt first
Every example shows the exact words used, so you can see which parts of the description did the work.
AI Music Examples
Eight prompts, taken apart. Each shows what comes back, which instruction decided it, and what happens to the output when that instruction is removed.
In short
These are worked examples rather than a gallery: each one shows a real prompt, describes what it produces, identifies the specific word that decides the result, and shows what changes when you take that word out. The pattern matters more than the prompt.
Remix the prompt — tap any highlighted part
A track at — . Instrumental, .
lo-fi hip hop, 78 BPM, warm and unhurried. Instrumental, vinyl crackle, highs rolled off.
01 · Documentary / voiceover
A bed that disappears under narration
The prompt
Restrained documentary underscore, 72 BPM, A minor. Felt piano, sustained cello, light shaker. Steady with no builds, instrumental, midrange kept sparse for voiceover.
What comes back
A slow, even piece with almost no dynamic movement. Piano sits low and soft, cello holds long notes underneath, and the shaker keeps time without drawing attention. Nothing arrives or resolves — it simply continues, which is what lets a voice sit on top of it for ten minutes without fatigue.
Why it works
“No builds” is the load-bearing instruction. Left out, almost every model adds a swell somewhere in the first minute, because build-and-release is what most training data does. “Midrange kept sparse” is the second: it tells the arrangement to leave 300 Hz to 3 kHz alone, which is where speech lives. Together they turn a piece of music into a bed.
Change one thing
Remove “no builds” — A swell appears around 40 seconds and pulls attention off the narration exactly when you least want it.
02 · Short-form video
A short loop that grabs in half a second
The prompt
Punchy 15-second loop, 145 BPM, half-time trap feel. Detuned bell melody, 808 slide on the downbeat, crisp hi-hat triplets, one vocal chop stab. Hook lands immediately, loops seamlessly.
What comes back
A dense, immediate fifteen seconds. The bell figure states itself in the first bar rather than building toward one, the 808 slides into the downbeat, and hi-hats fill the space between. It ends where it began, so the loop point is inaudible.
Why it works
“Hook lands immediately” inverts the default. Left to itself a model writes an intro, because songs have intros — and on short-form video an intro is the part nobody hears before scrolling. “Half-time” is the other key word: at 145 BPM without it you get something frantic; with it, the kick and snare land at half speed and the track feels heavy rather than fast.
Change one thing
Drop “half-time” — The same 145 BPM reads as drum and bass energy instead of trap weight — a completely different genre from one removed word.
03 · Streaming / study
Music that survives six hours
The prompt
Chilled lo-fi, 92 BPM, F major. Dusty swung drums slightly behind the beat, warm Rhodes chords, soft sub-bass, faint vinyl crackle, highs rolled off. One level throughout, instrumental, loops seamlessly.
What comes back
A warm, slightly blurred loop with no events in it. Drums drag a fraction behind the grid, the Rhodes cycles through four bars, and the vinyl noise sits underneath everything as texture. Nothing in it is memorable, which is the specification rather than a failure.
Why it works
“Slightly behind the beat” is what separates this from generic chill music. Perfectly quantised lo-fi sounds mechanical; the drag is the genre. “Highs rolled off” and “vinyl crackle” ask for the artefacts other genres remove — lo-fi is defined by what it takes away. “One level throughout” is the endurance instruction: anything with a peak becomes irritating by the third hour.
Change one thing
Add “with a melodic lead” — It becomes listenable rather than background — better for a first play, worse for the fifth hour of a stream.
04 · Advertising
Thirty seconds that end exactly on thirty
The prompt
Thirty-second advertising bed, 120 BPM, C major. Plucked synth, light claps, rising sweep at 22 seconds, resolved chord landing exactly at 30 with a clean stop, not a fade. Instrumental, midrange open for voiceover.
What comes back
A bright, uncluttered thirty seconds with a clear shape: steady for twenty, a rising sweep, then a chord that lands and stops. No fade, no tail — it finishes on the beat you bought.
Why it works
“Clean stop, not a fade” is the instruction that makes this usable. A fade means the edit has to cut into decaying audio, which is audible. Naming the sweep position (“at 22 seconds”) gives the arrangement a landmark to build toward, which is far more reliable than asking for “a build”. Generating at the target length beats trimming a longer track, because trimming cuts mid-phrase.
Change one thing
Ask for 15 seconds instead — The arrangement compresses rather than truncating — the sweep moves to around 10 seconds and still resolves on time.
05 · Songwriting
A song with a chorus that actually lifts
The prompt
Contemporary pop, 102 BPM, F# minor. Female vocal, conversational verses, pre-chorus lift, big open chorus with stacked harmonies. Piano and programmed drums in verses, guitars entering in the chorus. Two verses, chorus twice, bridge, final chorus.
What comes back
A structured song rather than a loop. Verses sit close and restrained with piano and programmed drums; the pre-chorus tightens; the chorus opens out with guitars and layered harmony. The bridge changes perspective once before the final chorus.
Why it works
The explicit structure — “two verses, chorus twice, bridge, final chorus” — is what stops the result being a section that repeats. The other decisive instruction is the instrumentation split: naming what enters in the chorus makes the chorus feel bigger through arrangement rather than volume, which is how the lift is produced in real records.
Change one thing
Remove “pre-chorus lift” — The chorus still arrives but lands flat — the pre-chorus is what creates the expectation the chorus pays off.
06 · Games
A loop you can hear two hundred times
The prompt
Fantasy game menu loop, 80 BPM, D dorian. Solo cello melody over harp arpeggios, soft choir pad. Patient and unhurried, instrumental, seamless loop with no obvious start point.
What comes back
A slow, circular piece that gives no clue where it began. The cello states a short phrase, the harp cycles underneath, and the choir pad holds the harmony without moving much. It can run for twenty minutes on a menu screen without asserting itself.
Why it works
“No obvious start point” is the instruction that makes it loop properly. A piece with a clear opening gesture announces its restart every cycle, and after fifty repeats that becomes the only thing a player hears. D dorian does quiet work too: the mode reads as old and slightly unresolved without being sad, which is why so much fantasy scoring lives there.
Change one thing
Change to D minor — The same arrangement turns melancholy. Dorian’s raised sixth is the difference between wistful and bleak.
07 · Podcast
Fifteen seconds that set expectations
The prompt
Podcast intro, 105 BPM, 15 seconds. Warm Rhodes chord, upright bass walk, brushed snare, one vibraphone flourish at the end. Conversational and inviting, instrumental, ends on a resolved chord.
What comes back
A short, warm figure with a clear ending. The Rhodes states a chord, the bass walks underneath, brushes keep time, and a single vibraphone phrase closes it. It resolves rather than fading, so the host can start speaking cleanly.
Why it works
“Ends on a resolved chord” is what makes it usable week after week — an intro that fades leaves the host guessing when to come in. “One vibraphone flourish” is deliberately singular: asking for a flourish gets you a flourish, asking for flourishes gets you a busy fifteen seconds that fights the first sentence of the episode.
Change one thing
Ask for 8 seconds — The bass walk disappears and the vibraphone lands sooner — tighter, and better for a show that opens cold.
08 · Hip hop
Bars that sit in the pocket
The prompt
Boom bap instrumental, 90 BPM, C minor. Hard sampled drums with slight swing, dusty jazz piano loop, upright bass. Midrange left open for a vocal. Instrumental.
What comes back
A four-bar loop with a heavy backbeat and a short piano figure that repeats. The swing pushes the hats slightly late. The middle of the mix is conspicuously empty — nothing competes with where a voice would sit.
Why it works
“Slight swing” is the difference between a beat that moves and one that merely plays; perfectly quantised boom bap sounds like a demo. “Midrange left open” is the instruction most people forget when generating an instrumental to rap over — without it, the piano fills exactly the frequencies the voice needs.
Change one thing
Raise to 140 BPM and ask for triplet hi-hats — Same instrumentation, different genre entirely — it becomes trap, and the pocket moves to half-time.
Weak prompts and what to write instead
The left column is what people actually type. None of it is wrong exactly — it is just under-specified, and the model fills the gaps with whatever is most common in its training data.
| Weak | Stronger | Why |
|---|---|---|
| chill music | lo-fi hip hop, 78 BPM, warm Rhodes, dusty swung drums, vinyl crackle, instrumental | The weak version names a mood with no acoustic content. Every parameter that decides the sound — tempo, instruments, texture — is left to chance. |
| epic cinematic music | cinematic orchestral, 90 BPM, D minor. Lone piano and low strings, taiko entering at 20s, full brass at 45s, hard stop at 60s | “Epic” describes an effect, not a cause. Naming where the build starts and where it lands is what produces the effect. |
| happy upbeat song for my video | bright indie-dance, 124 BPM, A major. Clean guitar riff, disco hi-hats, claps on 2 and 4. Instrumental, hook in the first two seconds | Two adjectives and a use case give the model nothing to act on. The strong version specifies the four things that actually change the output. |
| a rap beat | boom bap, 90 BPM, C minor. Hard sampled drums with slight swing, dusty jazz piano, upright bass, midrange open for vocals | Rap spans 85 to 150 BPM across subgenres that share almost nothing. The tempo and the drum treatment are the genre. |
| background music, not too loud | documentary underscore, 72 BPM, A minor. Felt piano and sustained cello, no builds, midrange sparse for voiceover | Loudness is a mix decision you make afterwards, not something a generator controls. What you actually want is flat dynamics and a clear midrange. |
| song like [famous artist] | contemporary R&B, 72 BPM, Bb minor, half-time. Smooth female lead with stacked harmonies, warm Rhodes, deep sub-bass, finger snaps | Naming an artist is unreliable and legally awkward. Describing what makes that artist sound the way they do is both more effective and safe to release. |
The pattern behind all of them
Every strong prompt above names the same five things: genre, tempo, instrumentation, an explicit decision about vocals, and one structural instruction — where a build lands, how it ends, what to leave out. The structural instruction is the one most people skip and the one that most often decides whether the result is usable, because it is the only part the model cannot guess from genre convention.
Note how often the decisive instruction is a negative — “no builds”, “no obvious start point”, “midrange left open”, “clean stop, not a fade”. Telling a model what not to do constrains it far more sharply than another adjective, because adjectives are things it can satisfy loosely and prohibitions are not.
Inside
What a sentence turns into
The prompt, and the track it produced.
1 / 5
You can hear the difference
A feature list tells you a tool supports twelve genres. An example tells you whether its lo-fi is any good.
The prompt is the lesson
Seeing the words that produced a result teaches more than any amount of description.
They set expectations honestly
Examples chosen to be representative rather than exceptional are more useful, even when less impressive.
They are a starting point
Every example is meant to be copied and altered, not admired.
Frequently asked questions
What do these examples show?
The exact prompt, what it produced, the specific word that decided the result, and what changes when you remove it. The pattern matters more than any individual prompt.
Can I copy these prompts?
That is what they are for. Copy one, change a single thing, and compare — changing one variable at a time is the fastest way to learn what each word does.
Why does one word change the result so much?
Because most prompt words are vague and a few are load-bearing. Tempo, instrument names and production era are levers; adjectives like "nice" or "epic" are resolved by the model rather than by you.
Do these prompts work on other AI music generators?
Mostly. Genre, tempo and instrumentation transfer well across models. Structure tags and negative instructions are the parts that behave differently platform to platform.
What is the most common prompting mistake?
Describing a feeling instead of a sound. "Emotional" produces an average of everything the model has seen labelled that way; "solo cello, no percussion, 60 BPM" produces a specific record.
How many examples should I work through?
Three or four is enough to see the pattern. The point is not to collect prompts but to notice which kinds of word change a result and which are decoration.
AI Music Examples
Take a prompt, change one thing, and hear the difference.