Song Generator Prompts: Genre, Mood and Tempo

Short answer: a song generator prompt works best as a short ordered brief rather than a sentence. Name the genre family first, then the mood, then the tempo, then the instrumentation, then the voice, and only then the structure. Each line narrows the space the next one works in, so the same six decisions produce a far more usable track when they are made in that order than when they are bundled into one paragraph.

This is what each part of a prompt actually controls, the words that carry signal and the ones that do not, and a repeatable way to iterate on a song generator without losing the version you liked.

What a song generator prompt actually controls

Six decisions do almost all of the work. Everything else is refinement, and adding more detail before these six are settled usually makes the output less predictable rather than more.

Genre family and era. Start at the family level — folk, house, boom-bap hip hop, synth-pop — then add an era if it matters (late-70s disco rather than disco). Family sets the tempo range, the instrumentation and the production conventions all at once.
Mood. The emotional target, not the genre: warm and reflective, tense, celebratory, wistful. Mood changes how the same chord progression is voiced and how the vocal is delivered.
Tempo and feel. Give a number in BPM plus a word for the feel (driving, laid back, half-time). A number alone can still be interpreted as a slow song with busy percussion.
Instrumentation. The two or three sounds that should be recognisable in the mix, and the one that should lead. Naming eight instruments rarely produces eight audible instruments; it produces a wash.
Voice. Whether there is a lead vocal at all, its register and character, whether harmonies or backing layers exist, and whether the delivery is intimate or full-throated. Vague terms like “good singer” describe nothing the generator can act on.
Structure and length. If the shape matters — short intro, two verses, a bridge, an outro that fades — say so, because structure is the part people most often notice after generation rather than before.

The order to write it in

Write the prompt as separate lines in the order above. Each line constrains the next, so you are narrowing a search rather than describing a wish, and when the result is wrong you can see which line caused it.

Line 1 — genre family and era. Anchor everything else.
Line 2 — mood. Sets intensity and colour.
Line 3 — tempo in BPM plus feel. Fixes the groove.
Line 4 — two or three instruments, one lead. Fixes the texture.
Line 5 — voice: present or absent, register, delivery, harmonies. Fixes the front of the mix.
Line 6 — structure and length. Fixes the shape.

A prompt written this way also reads as documentation. The line that says 95 BPM, brushed drums, upright bass is a precise instruction you can reuse and modify; “a cool retro song” is not, because nothing in it can be varied while everything else stays fixed.

Vague versus specific: what changes in the output

The difference between a prompt that produces something usable and one that produces something forgettable is usually not length, it is whether each word is a decision.

Vague: “an emotional pop song with good production and a nice voice.” Problem: “emotional”, “good” and “nice” are judgements, not parameters, so the generator fills the gaps with its own defaults.
Specific: “synth-pop, melancholic but hopeful, 108 BPM with a steady four-on-the-floor pulse, warm analogue pad and a bright plucked lead, female lead vocal close-miked with a quiet octave harmony in the chorus, verse–chorus–verse–chorus–outro.” Result: every phrase is something the model can choose against.
Overloaded: “orchestral trap ballad with distorted 808s, saxophone solo, choir, banjo, half-time and double-time drums, whispered and shouted vocals.” Problem: contradictions do not add richness, they add coin-flips, and you cannot tell which one the next generation will resolve.

Change one variable at a time

Iteration is where most of the quality comes from, and the failure mode is changing everything at once, then having no idea which change helped. Keep the six lines in a text file, hold five of them fixed, and edit exactly one per generation. When a result is close, change the smallest line available — tempo by five BPM, the lead instrument, or the vocal delivery — rather than rewriting the brief. Save the prompt that produced anything you would keep, because a song generator is not reproducible from memory: the version you liked is a version you should have written down.

Where lyrics fit, and where they do not

Separate the words from the sound. Decide the structure, registers and subject of the lyrics first, then let the prompt describe how they are sung. Mixing a full lyric sheet into a sound description usually muddles both, and rewriting lyrics inside a prompt you have already tuned throws away the tuning. EchoLora’s guide to writing lyrics for a song generator covers line length, syllable stress and the chorus/verse contrast that carries the hook, while the four-step AI music workflow shows where prompting sits in the wider process from idea to finished track.

Check the track before it leaves the session

Listen for structure first, mix second, lyric intelligibility third. A generator can deliver a strong chorus and a verse that drifts out of key, or a vocal that is burying the one line the song depends on. Before publishing anywhere, check the platform’s current terms on how generated audio may be used commercially — the licensing position for AI-assisted music is worth reading directly rather than assuming, and any claim about commercial use should come from those terms, not from a forum post. How to review a generated track sets out what to listen for in order.

Frequently asked questions

How long should a song generator prompt be?

Long enough to state six decisions and no longer. Two to four sentences, one per decision, is usually the sweet spot. Extra adjectives that describe quality rather than sound — good, professional, catchy — add length without adding information, and contradictory instructions add unpredictability rather than range.

Should I name artists or existing songs in a prompt?

Describe the sound instead. A named artist is a request the platform may decline, and even when it is accepted it gives you a result you cannot describe in your own words or reproduce deliberately. “Warm tape-saturated drums, fingerpicked acoustic guitar, close vocal” is not just safer, it is a reusable instruction you can tune line by line.

Why do I get different results from the same prompt?

Because generation samples from a space rather than replaying a file. Treat the first result as one draw, not the answer: run the same prompt two or three times to see what is stable, then vary one line at a time. If a prompt only works once, it was luck; if it works three times, it is a specification.

Can I control tempo, mood and instrumentation at the same time?

Yes, and that combination is exactly what a well-ordered prompt is for. The reason to keep the lines separate is diagnosis: when a track comes back wrong you can point at mood or at tempo rather than guessing, and your next change stays small enough to judge.

Do I need to write lyrics before or after generating the music?

Before, if the song is meant to carry a message. Settle the structure and the emotional arc of the words, then prompt for the sound that carries them. Writing lyrics after generation tends to force the words into a melody that was shaped for something else.

Put the prompt to work

EchoLora is built around this order of decisions: choose the style, describe the sound, set the structure, then generate and iterate on the track. Browse the catalog to hear how a style lands before you write your own prompt, and if you are publishing your results, the creators section covers sharing what you make.