MuseGen MuseGen Team
9/14/2026
The formula is the easy part. Most people can name the slots inside a minute — genre, tempo, instruments, vocal — and then stall in the same place, cursor blinking, because knowing a slot exists is not the same as knowing what word goes in it. The template below takes ninety seconds to learn. The rest of this page is the vocabulary, laid out to be looked up rather than read.
In short #
One sentence, six decisions, roughly in this order. Fill what you have an opinion about and skip the rest on purpose, knowing that a skipped slot is still answered by the tool. The tables further down are the word list for each slot.
- Genre, with one qualifier. A broad genre word is a bundle of averages. The qualifier is the part that changes the output.
- Tempo, and the feel that goes with it. A number, plus how the beat is subdivided. Straight and swung at 96 BPM are different songs.
- A lead, and a job for everything else. Who carries the melody, and what the other parts are doing while it does.
- The vocal, cast or switched off. Range, delivery and stacking, or the single word instrumental.
- Production and era. How the recording sounds: space, texture, decade. Not how it makes you feel.
- Length, and how it ends. A duration and a final gesture, because the default everywhere is a fade.
Quick facts #
- The template: Genre, tempo and feel, lead and roles, vocal, production, length and ending — in that order.
- Most skipped slot: The ending. It is also the slot that most often ruins an otherwise usable take.
- Mood words: Not a slot of their own. Translate them into the other six before typing them.
- Tempo: A number, not an adjective. "Fast" spans about forty BPM and shifts by genre.
- Words per slot: Three to ten. Six filled slots lands somewhere near forty-five words.
- How to use this page: As a lookup. Open the slot you are stuck on and take the right-hand column.
The template, in one line #
Here is the whole formula. It fits on one line because it is not really a formula, in the sense of something you calculate; it is a list of six questions asked in a sensible order, with the load-bearing ones first.
The template
[genre + qualifier], [BPM] and [feel]. [Lead] carries the melody, [other parts and their jobs]. [Vocal, or instrumental]. [Production and era]. [Length], [ending]. Filled in
country, 96 BPM with a straight backbeat. Pedal steel carries the melody, acoustic guitar and brushed snare keep time, fiddle answers at the end of each line. One male lead, plain and unhurried, no harmony stack. Dry and close, 1970s studio sound. 2 minutes 30, ending on a single strummed chord.
Fifty words, six slots, nothing decorative. Every phrase names something a musician in a room would physically do differently on hearing it, which is the only test that matters when you are deciding whether a word has earned its place.
That is the formula section over, deliberately. The order earns a sentence — front-loaded clauses carry more weight in most tools, so genre and tempo go first and the ending last — and the reasoning behind each slot is covered at length in our eight rules for writing AI music prompts. The film-scoring variant, built around scene and picture, lives in the cinematic prompt guide.
What neither of those pages does, and what this one is for, is answer the question that actually stops people: fine, but what do I type? Everything below is a lookup table — left column the thing you were about to write, right column what to write instead. Not exhaustive, and not rules: just the translations that save you the twenty generations it takes to find them yourself.
https://www.youtube.com/watch?v=M0wOyr5cM5Y A short, medium and detailed prompt compared inside one specific tool, where the template on this page is written to be tool-agnostic.
Slot 1: how specific the genre should be #
A genre word is a compression format. Writing rock does not select a sound so much as hand over a large bundle of defaults, which the model unpacks by reaching for whatever sits nearest the middle. That is why broad genre words come back technically correct and completely characterless: you asked for the average and got it.
The fix is not more genre words. It is one, narrowed once. A single qualifier — a decade, a region, a subgenre, a defining instrument — cuts the bundle down to something with a shape. Two is the working ceiling. A third starts arguing with the first two, and what comes back is a blend of all three rather than any of them.
| Too broad | One qualifier that changes the output |
|---|---|
| pop | 1980s synth-pop, gated snare and bell-like digital keys |
| rock | 1990s alternative rock, quiet verse into a loud chorus |
| electronic | UK garage, shuffled hats over a swung two-step beat |
| jazz | hard bop quintet, trumpet and tenor trading choruses |
| folk | Appalachian old-time, fiddle and clawhammer banjo |
| metal | doom metal, half-time riffs tuned well down |
| latin | cumbia, accordion lead over a shuffling guiro pattern |
| hip-hop | boom bap, a sampled loop and swung sixteenth-note hats |
| ambient | drone ambient, one sustained chord and no percussion |
The boundary: qualifiers work because they are subtractive. Each rules something out, which is why one good qualifier beats four genre names stacked with plus signs. If you want three genres at once, pick the one that decides the drums and demote the others to a single named element borrowed from each — an instrument, a production trait.
Slot 2: tempo, and the feel words #
Tempo is the cheapest specific thing in the template, and it is half a decision. The number says how fast the pulse is. It says nothing about how that pulse is divided, and division is where most of what people call groove actually lives. Ninety-six BPM straight and ninety-six BPM swung are not variations of one another; they are two different pieces of music sharing a metronome setting.
So this slot wants two things: a number, and a feel word behind it. The words below are the ones that reliably do something, and several are worth knowing precisely, because they are commonly used to mean the opposite of what they mean.
| Feel word | What it actually asks for |
|---|---|
| straight | Even subdivisions, no swing. The default in most pop, rock and techno. |
| swung | The offbeat lands late. Native to jazz, blues, boom bap and UK garage. |
| half-time | The backbeat moves to beat three, so 140 BPM feels like 70 without slowing down. |
| double-time | Subdivisions halve. The track feels twice as fast at the same BPM. |
| four-on-the-floor | Kick on every beat. House, techno, disco, most dance music. |
| shuffled or triplet | Divided in threes rather than twos. A rolling, uneven pulse. |
| laid-back | Snare and bass sit a fraction late. Reads as relaxed rather than slow. |
| pushed or driving | Hats and snare sit slightly early. Urgency without adding BPM. |
| syncopated | Accents land off the main beats rather than on them. |
| free or rubato | No fixed grid at all. Followed inconsistently, so say it only if you mean it. |
The boundary: if you are chasing a reference track, do not estimate its tempo. Estimates run low, and ten BPM out changes the genre the model thinks you asked for. Our guide to what BPM is has typical ranges by genre, and BPM detection reads the real number off an audio file.
The missing slot: what to write instead of a mood word #
This is the section people arrive for, and it is the one place the template has no slot at all. Mood is not a variable. Chill, epic, dreamy and gritty describe what happens in a listener when the music works, not what the music does, so there is nothing in them for a model to act on — and because everyone reaches for the same dozen, prompts built out of mood words converge on the same handful of outputs.
The answer is not to stop thinking in moods. Moods are how anyone with ears decides what they want. It is to translate the mood into the other six slots before it reaches the box, which takes about two seconds once you have seen it done.
The same intent, written twice — the right-hand column is the version a model can act on.
| You meant | Write this instead |
|---|---|
| chill | slow tempo, brushed or soft drums, sparse arrangement, no loud cymbals |
| vibey | a short loop repeated with little variation, one sustained pad underneath |
| epic | wide dynamics: a quiet opening, then full drums, low brass and choir |
| dark | minor key, low register, slow-attack pads, nothing bright on top |
| warm | tape saturation, rolled-off treble, upright bass or electric piano rather than synth |
| dreamy | long reverb, slow chord changes, soft attacks and no percussive transients |
| gritty | overdriven bass or distorted guitar, drums pushed until they clip |
| uplifting | major key, a rising melodic line, a chorus that adds instruments rather than volume |
| nostalgic | name the era instead: 1960s tape, 1980s FM keys, VHS-era synth pads |
| ethereal | high sustained voices, no low end, long tails, no drum kit |
| aggressive | fast, loud, distorted, short gaps between phrases, hard consonants in the vocal |
| sad | slow, minor, sparse, instruments entering one at a time rather than together |
| cinematic | name the scene and the forces: solo cello over low strings, no drums until the last third |
The boundary: you can keep one mood word, and often should. Placed after the concrete clauses it costs a single word and breaks ties on everything the other slots left open — which register the pad sits in, how hard the drums hit. What it cannot do is carry a prompt alone, which is the most common way prompts go wrong.
Slot 3: giving every instrument a job #
This is the slot people over-fill and under-specify at the same time. Naming eight instruments feels like being detailed. What it communicates is that eight instruments exist, and with no other information the safest thing a model can do with eight instruments is have all of them play, all the way through. That is the wall of sound so many generations come back as.
The vocabulary this slot needs is not more instrument names. It is role phrases: short verbs saying what a part is doing relative to the others. Two or three turn a list into an arrangement, at about four words each.
| Role phrase | What it tells the arrangement |
|---|---|
| carries the melody | This is the lead. Everything else is support and should stay out of its way. |
| answers between phrases | Fills the gaps the lead leaves, and never plays over the top of it. |
| keeps time | Steady and unvarying, sitting behind everything else in the mix. |
| enters only on the chorus | Silent everywhere else, which is what makes the chorus feel like a lift. |
| drops out at the bridge | An exit. Removing a part is as audible an event as adding one. |
| doubles the vocal | Plays the sung line in unison or an octave away, thickening it. |
| sits behind the beat | Slightly late, deliberately. The main lever for a relaxed feel. |
| pads underneath | Sustained chords with no rhythmic role, filling the space between parts. |
| trades with the lead | Alternating phrases, call and response, neither one dominant. |
| one hit per bar | Punctuation rather than a part. Keeps an instrument present without crowding. |
The boundary: three roles is plenty, and the lead is the one that must be named. Beyond three, the phrases compete for the same space and something gets dropped, usually whatever you wrote last. For entrances and exits in more detail than a prompt allows, our piece on what arrangement is covers the same ideas at length.
Slot 4: casting the vocal #
Leave this one open and nearly every tool reaches for the same thing: a clean, mid-range, lightly processed pop lead, front and centre. It is a perfectly good voice. It is also the voice on several million other generations, which is why this slot decides more than any other whether a track sounds like yours.
Casting one takes about eight words. The dimensions below run roughly in order of how much difference they make, and three at once is plenty.
| Dimension | Words that work |
|---|---|
| Range | soprano, alto, tenor, baritone, bass — or plainly, low female voice, high male voice |
| Delivery | spoken, whispered, breathy, conversational, belted, shouted, half-sung |
| Phrasing | legato and sustained, clipped and staccato, rapid-fire, drawn out behind the beat |
| Stacking | one lead only, doubled, three-part harmony, full choir on the chorus |
| Backing | call and response, oohs and ahs under the verse, or no backing vocals at all |
| Distance | close to the mic and dry, or distant and washed in reverb |
| Restraint | no ad-libs, no runs, no melisma — three exclusions that change a lot |
| Absence | instrumental, or no vocals. Two words that save an entire generation. |
The boundary: this slot is how the voice sounds, not what it says. Lyrics belong in their own field; pasted into a style box they get read as description of the music instead, which is how you end up with a track about the concept of a chorus. Write them separately, or start from a lyrics generator. If the processing words above are unfamiliar, what a vocal chain is defines them.
Slot 5: production and era words #
Production words are the useful cousins of mood words. Warm, dry and wide look like adjectives from the same family as chill and epic, but they behave completely differently, because each describes a decision an engineer made in a room. There is a physical thing behind dry. There is nothing physical behind vibey. That is why this slot can be filled with adjectives when the mood section above could not.
The translations below are conventions, not specifications. Nobody publishes what these words map to internally, but their meanings are stable enough across engineers to survive the trip into a prompt intact.
| Production word | What it translates to |
|---|---|
| warm | Tape or valve saturation, a softened top end, low mids left in. |
| lo-fi | Audible noise floor, limited bandwidth, pitch instability. A whole genre is built on it. |
| punchy | Short, hard transients and a tight low end that gets out of the way. |
| wide | Stereo spread: doubled guitars, spread keys, bass and kick kept in the centre. |
| dry | Little or no reverb. Everything sounds close and immediate. |
| roomy | Audible reflections, as if the whole band was recorded in one space. |
| vintage | Name the decade instead, and let the gear cliches of that decade follow. |
| polished | Even levels, no rough edges, a modern commercial mix. |
| raw | One take, minimal editing, small mistakes left where they fell. |
| saturated | Harmonic distortion stopping short of obvious fuzz. |
| sparse | Fewer parts and more silence between them. |
| dense | Many layers with little space, everything overlapping. |
The boundary: these words describe the sound you want generated, not a process you run afterwards. A text-to-music tool renders a finished mix in one pass, so asking for a warm mix is a request about what gets made. Two production words is usually the right number. Four pull against each other, and wide, dry, dense and raw in one breath describes no recording that has ever existed.
Slot 6: length, and how it ends #
The quick facts above call this the most skipped slot, and it gets skipped because it reads as housekeeping rather than music. Left open, nearly every generator resolves one of two ways: a fade, or a stop at whatever second the clock ran out on. Both are fine when nobody is listening for them, and useless the moment the track has a job to do. Two short things go here — a duration, and a final gesture worth about four words.
| Ending | What it asks for |
|---|---|
| hard stop, no fade | Full band, one last beat, then silence. Abrupt on purpose. |
| ends on a sustained final chord | Everything lands together and rings out. The safest definite ending. |
| ends on the last vocal line | Instruments clear out first and the voice finishes alone. |
| strips back to one instrument | Parts leave one at a time until a single one is still playing. |
| fades over the last eight bars | A fade you asked for, with a stated length, rather than the default one. |
| ends on a cymbal swell | A rise into silence rather than a landing on it. |
| loops cleanly, no ending | The last bar leads back into the first. For background and game use. |
The boundary: length is a request rather than a setting. Most tools generate in fixed blocks and land near your number instead of on it, so treat two minutes thirty as a target. The ending is the more reliable of the two, which is the reverse of what most people assume.
Three finished prompts, short to long #
The tables are only useful assembled, so here are three, built from the columns above and deliberately at different lengths. None is the correct one. They fill four, five and six slots, and the trade is what you would expect: more slots filled, fewer surprises, in both directions.
Short · four slots
dub, 72 BPM, half-time. Bass carries the melody, guitar skank on the offbeat drenched in tape delay. Instrumental.
Medium · five slots
shoegaze, 132 BPM, straight eighths. Two layers of distorted guitar carry the melody, bass keeps time, drums pushed until they clip. One breathy female lead, doubled and buried well under the guitars. Dense and washed. 3 minutes 30.
Long · six slots
salsa dura, 96 BPM, syncopated montuno feel. Piano carries the montuno, a full brass section answers between phrases, congas and timbales keep time, bass drops out at the bridge. One male lead, conversational in the verse and belted on the chorus, with a three-part coro answering. Dry and close, 1970s New York studio sound. 4 minutes, ending on a full-band stop.
Nineteen words, thirty-eight, sixty. The short one is not worse; it hands three decisions to the tool on purpose. What it cannot do is complain about how those three came out, since nobody asked for anything in particular.
Read them back against the six slots and you can see where each stops. The dub prompt names a genre, a tempo with a feel, two roles and the absence of a vocal, then goes quiet: production and length are left open, which for dub is a reasonable bet, since the genre word already implies most of the production. The shoegaze prompt adds a vocal and a length but skips the ending. The salsa prompt fills everything, and at sixty words sits at the top of the useful range rather than past it.
Which is the one thing these tables can lead you past. Having a vocabulary for every slot does not mean using all of it. Six slots at ten words each is sixty, and beyond roughly that point clauses get diluted rather than obeyed, the later ones first. A prompt running long is almost never too many slots. It is one slot with an unroled instrument list in it.
What the template is actually for. It does not write the song and it does not raise the ceiling — a filled-in prompt can still come back dull. What it does is stop you handing over decisions by accident, which is the difference between a track that is yours and one that is the average of everything the model has heard. The tables are meant to be reopened, not memorised.
Where this vocabulary goes in MuseGen #
Half of this page already exists as a set of dropdowns in our own tool, which turns out to be a fair way of showing both what a word bank is good for and exactly where one runs out.
Our own AI song maker takes a written description and renders a full track from it. The style box comes with Genre, Moods, Voices and Tempos pickers — four of the six slots as lists, which is convenient and also the ceiling of the idea: a dropdown can hand you dreamy, but the clause dreamy translates into is something you type. So the pattern that works is pickers for the coarse choices, free text for the right-hand columns above: the qualifier on the genre, the feel word behind the tempo, the roles, the ending. The hint under the box is honest about the rest, saying plainly that anything left blank is filled at random. Lyrics get a field of their own, and BPM detection gives slot two a real number instead of a guess.
Four of the six slots as dropdowns — the qualifiers go in the free text beside them.
Slot four splits in two: how the voice sounds is a prompt, what it sings is not.
Nothing in these tables is a switch, in our tool or in any other. The same filled-in prompt run twice comes back as two different songs, and a clause from the right-hand column is sometimes honoured loosely. What a full template buys is a narrower set of outcomes to choose between, not a described result delivered on request.
FAQ #
Is there one prompt formula that works for every genre?
The six slots hold up across genres because they are the decisions every recording makes: what style, how fast, who leads, whether anyone sings, how it was recorded, how long it runs. What changes is which slot carries the weight. In electronic music the tempo and production slots do most of the work. In folk or jazz it is usually the lead and roles slot. The template stays the same; the amount you write in each slot does not.
Do I have to fill in every slot?
No, and filling all six is not automatically better. Every slot you leave blank still gets an answer, chosen from whatever the model does most often, so the only question is whether you care about that particular decision. Leaving melody and chord voicings open is usually deliberate. Leaving the vocal and the ending open is usually an accident, and those two are where most generic results come from.
What is the difference between a mood word and a production word?
A mood word describes how a track lands on a listener: chill, epic, dreamy. A production word describes how the recording was made: dry, saturated, wide, lo-fi. The second kind is a statement about the audio, so it survives being turned into sound. The first kind has to be translated into tempo, instrumentation, register and arrangement before it means anything at all. Translating it is what the tables on this page are for.
Can I paste a list of words instead of a sentence?
You can. Short comma-separated clauses work fine, and most style boxes read description rather than prose. What does not work is a bare pile of nouns with no relationships, because eight instruments with no roles tells the model that all eight play all the time. Keep the list if you prefer it, but attach a verb to the parts that matter: which one leads, which one answers, which one appears only in the chorus.
Should tempo be a number or a word?
A number, then a word. Fast covers a forty-BPM range and means something different in drum and bass than it does in country, while a number lands close in most tools. The word that follows it is the feel: straight, swung, half-time, four-on-the-floor. If you are matching an existing track, measure its tempo rather than estimating it, because guesses tend to land low.
What do I write if I do not know the genre?
Describe the instrumentation and the feel, and let the genre slot stay empty. Nylon-string guitar, upright bass and brushed drums at 74 BPM is a more useful instruction than a genre name you are unsure of, because it names actual sound sources rather than a category. A wrong genre word is worse than no genre word, since it pulls the whole arrangement toward a bundle of defaults you did not want.
Does the same template work for instrumental tracks?
Yes, with slot four reduced to a single word. Write instrumental, or no vocals, and the attention that would have gone into casting a singer moves to the lead and roles slot, which matters more when there is no voice on top. Instrumentals live or die on arrangement movement, so entrances and exits are the part worth being specific about.
Can I reuse one prompt across different tools?
The vocabulary transfers. The parsing does not, entirely. Genre names, tempo numbers, instrument roles and exclusions are read similarly across most text-to-music tools, so the same sentence is a reasonable starting point anywhere. Where tools differ is in field layout, how strictly they honour exclusions, and how much text they read before later clauses get diluted. Expect to re-tune the length rather than rewrite the content.
Keep reading #
- [How to Write AI Music Prompts: 8 Rules for Better Results](/blog/how-to-write-ai-music-prompts) — the reasoning behind each of the six slots.
- [How to Write Perfect AI Music Prompts for Cinematic Tracks](/blog/how-to-write-perfect-ai-music-prompts-for-cinematic-tracks) — the same template, narrowed to scoring for picture.
- [What Is BPM?](/blog/what-is-bpm) — the numbers behind slot two, with typical ranges by genre.
- [What Is Arrangement in Music?](/blog/what-is-arrangement-in-music) — where the role phrases in slot three come from.
Sources #
- Video: "3 Levels of Prompting in Suno Ai" — youtube.com
- Field layouts, prompt length ceilings and how strictly exclusions are honoured differ from one generator to the next, and shift again with each model release. Whichever one you are using — suno.com ,udio.com or another — its own docs are the only current authority on how it reads a prompt.
- Related vocabulary in longer form: What Is Arrangement in Music? ,What Is a Vocal Chain? andWhat Is a Chord Progression? .