Writing lyrics and structure the model can actually sing

The model reads shape before meaning

It's tempting to treat lyric input as poetry and assume the model will find the music in it. In practice, the structural signals — where sections start, how long lines are, how many stressed syllables sit in a line — drive the result far more than the imagery does.

That's not a limitation to resent; it's the same constraint every songwriter works under. A line that doesn't scan doesn't sing, whether a human or a model is performing it. Writing for the model is mostly just writing for singability, made explicit.

This has a liberating implication. You do not need to write better poetry to get better results; you need to write more singable lines. Plain vocabulary in a clean rhythmic shape outperforms dense, clever imagery in an awkward one almost every time, and that's true of human singers too.

Use section tags, always

Untagged lyrics force the model to guess the form, and it guesses conservatively — usually a flat verse-chorus with no dynamics. Tags cost nothing and immediately improve arrangement.

• Tag every section, including instrumental ones like [Intro] and [Outro].
• Reuse the exact same chorus text each time you want the same melody — variation invites a new melody.
• Mark instrumental passages explicitly, e.g. [Guitar Solo], rather than leaving a gap.
• Keep tags on their own line with nothing else beside them.

[Intro]
[Verse 1]
[Pre-Chorus]
[Chorus]
[Verse 2]
[Chorus]
[Bridge]
[Chorus]
[Outro]

Line length and syllable discipline

Habits that ruin a generation

• Very long lines — over about fifteen syllables the delivery becomes a mumble.
• Parenthetical stage directions inside lyric lines; they often get sung.
• Inconsistent chorus text between repeats, which fragments the hook.
• Dense internal rhyme in slow ballads, which forces an unnatural rhythm.
• Unusual proper nouns and invented words — expect mispronunciation and plan for it.
• Numerals and symbols; write 'nineteen eighty four', not '1984'.

A verse that works, and why

Eight to nine syllables per verse line, consistent stress pattern, plain vocabulary, hard end rhymes. The chorus drops to four and six syllables with a repeated opening phrase — that contrast in density is what makes a chorus feel like a chorus, and it works whether the performer is human or not.

Notice also that there is no imagery here doing heavy lifting. A kitchen light, a coat, some cars. The emotional content comes from the situation and the restraint, not from vocabulary — which is exactly the kind of writing that survives being sung by a voice you did not choose.

[Verse 1]
The kitchen light is on at four
You left your coat beside the door
I count the cars that never stop
And wait for something else to drop

[Chorus]
Stay a while
The morning's slow
Stay a while
I'm not ready to let you go

Iterating lyrics without breaking everything

• Lock the style prompt completely while you edit lyrics.
• Change one section per run, not the whole lyric.
• Read every edit aloud at the target tempo before generating — most problems are audible in your own voice first.
• If a line consistently gets rushed, cut two syllables rather than slowing the tempo.
• Keep a copy of the version that scanned best; edits often make things worse.
• Once the lyric sings well, freeze it and go back to tuning style.

FAQ