How to Write a Music Generation Prompt That Lands

Briefing a music generator is not like briefing an image generator, and most of the frustration comes from assuming it is. A picture has no controls outside the sentence, so people learn to stuff everything into the sentence. A track has controls — key, tempo, meter, length, vocals — and the sentence is only responsible for the part those controls can't express.

Get that division right and most prompt problems disappear. Here's the division, and the handful of habits that come with it.

Rule one: the brief describes style, nothing else

The brief is for genre, instruments, mood, timbre, era and production. All six of those land:

warm lo-fi with electric piano, soft drums, late-night mood

80s synthwave, gated snare, wide analogue pads, tape saturation

solo nylon guitar, close-mic'd, room tone, unhurried

What does not belong there is anything with its own control. Tempo and key have fields, and those fields are sent as instructions the generator has to honour. Writing "120 BPM, in A minor" into the sentence as well doesn't reinforce them — it just gives the generator a second, weaker copy of the same request, and if the two ever disagree, you have briefed against yourself. "Slow ballad" over a fast pinned tempo is the classic version of this, and it produces exactly the confused result it sounds like.

Say the mood in words. Say the numbers in the numbers.

Rule two: detail decides freedom

This is the single most useful thing to internalise, and it isn't a quality setting — it's a consequence of how much you said.

• Write less and the generator invents more. "dreamy ambient" is an invitation. You'll get variety across takes and occasionally something better than you'd have thought of.
• Write more and it stays close. "dreamy ambient, one detuned Rhodes, no percussion, long reverb tails, single sustained bass note" is a specification. Takes will resemble each other and resemble your sentence.

Neither is the correct amount. They're two different jobs. Exploring? Say less. Filling a hole in an edit that has to be a particular shape? Say more. The mistake is writing a paragraph when you're exploring, or two words when you already know exactly what you need.

Rule three: instruments are the strongest lever

Of everything you can put in a brief, naming instruments changes the result most reliably. Genres are broad and mood words are subjective, but "solo cello", "brushed kit", "detuned Rhodes", "low brass" are unambiguous asks about what makes the sound.

Production language works the same way and is under-used: "close-mic'd", "tape saturation", "gated", "room tone", "lo-fi", "wide", "dry" all describe the recording rather than the notes, and they steer a take somewhere a genre label can't.

Rule four: chords are the one musical fact with no control

Key, tempo, meter and length are instructions. A chord progression is not — there's no separate input for harmony, so a progression travels inside the brief as words. It usually lands, and it is not guaranteed.

The studio is explicit about this: the fields are labelled "exact" and the progression is labelled "aim". That distinction is worth respecting in how you plan. If a progression has to be right — because a musician will play over it, or because it has to match an existing piece — generate a few takes and check the result rather than assuming.

Write progressions the plain way people write them, with the chord names a keyboard prints: Am F C G, not a theoretically immaculate spelling full of double flats. Common spellings are the ones a model has seen most.

If you're not sure what progression you want in the first place, the chord progression basics post covers how they're built, and the chord progression generator will lay out the diatonic options in any key.

Rule five: structure is written as tags, not described

Telling the generator "start quiet, then bring in drums, then a big finish" is a sentence about structure, and sentences about structure are weak. There is a stronger form: section tags, in order.

[Intro - ambient]

[Main Theme - piano]

[Climax - powerful]

[Outro - fade out]

The tags are the timeline. [Intro], [Verse], [Chorus], [Breakdown], [Guitar Solo], [Outro], [Fade Out] all carry a track's shape, and one dash-descriptor per tag steers the character of that section. This is how you say when an instrument arrives, which the brief alone can't express.

Two things to avoid: stacking descriptors ([Chorus - anthemic - huge - distorted] reads as words to be sung rather than as direction), and writing tags you don't mean. Leave the field empty and the generator shapes the track itself, which is often the right call.

Rule six: instrumental is a setting

Asking for "no vocals" in the brief barely works. The brief describes style, and a style description is a poor tool for suppressing an entire performance layer. The vocals control is what decides it, and it decides it properly.

The same is true in the other direction: if you want lyrics, they go in the lyrics field, not in the brief. And lyrics have their own craft — lines of roughly six to ten syllables, because syllables get matched to beats; CAPS sing harder; (parentheses) become backing vocals; a blank line separates sections.

Rule seven: know what the reading setting actually does

There are three readings of your brief. Word for word sends your text untouched. The other two let the generator re-draft it into a fuller description before rendering — lightly, or freely.

That's a real mechanical difference, and it's the only claim worth making about it. We ran our own comparison to find out whether a looser reading lands a requested key more often than a literal one. It couldn't tell them apart, in either direction. So choose on behaviour, not on a supposed accuracy ladder: pick word for word when you want your exact phrasing to be what's rendered, and a looser reading when you'd like your two-line idea fleshed out into something fuller.

One interaction is worth knowing because it surprises people: if you pin key, tempo, meter and length, there is nothing left for the generator to fill in, so it takes your brief as typed no matter which reading is selected. Leave one of the four on auto if you actually want a re-drafted brief.

Rule eight: repeatability has conditions

"Generate that again" is a reasonable thing to want, and it's possible — but only when a few things line up, because the re-drafting stage is the part that can't be pinned. A repeat needs the brief to go in word for word, the length to be fixed rather than automatic, and the planning stage to be off.

The trade is real: planning is what sketches structure, tempo and harmony before rendering, and it's what makes a pinned key and tempo actually hold. Turning it off buys repeatability and spends some of that reliability. The studio names whichever condition is still blocking a repeat right beside the control, so you don't have to remember this list.

Rule nine: judge in batches

A single take tells you about a single roll. Rendering several from the same brief tells you about the brief. If they come back alike, your description was tight; if they come back scattered, it was looser than you thought — and the axis they scatter along is exactly the thing to pin down next.

This is the fastest debugging loop available. "It ignored me" is nearly always "I said less than I thought I said".

A worked example

Start vague:

cinematic

That's an invitation to invent, and it will. Now specify the sound:

cinematic, low brass, solo cello, sparse percussion

Now the production:

cinematic, low brass, solo cello, sparse percussion, wide hall reverb, restrained

Now move the musical facts out of the sentence and into the controls — key, tempo, length, meter — and add structure as tags:

[Intro - solo cello]

[Build - low brass enters]

[Climax - full]

[Outro - cello alone]

Four small moves and you've gone from a mood word to a brief that describes a piece. Nothing here is a trick; it's just putting each kind of information where it's actually read.

The short version

• Brief = style. Fields = musical facts. Never both.
• Detail decides freedom, and both directions are useful.
• Instruments and production words steer harder than genre labels.
• Chords are an aim, not an instruction. Check them if they matter.
• Structure is tags in order, one descriptor each.
• Instrumental is a setting.
• The reading setting changes how much of your text is sent, and that's all we can honestly tell you about it.
• Render several. The spread is the feedback.

If you want the other half of the picture — which formats and scenes this is useful for, and how to fit a track to an edit — that's in background music for videos and games.