How it works

A quick tour of what the generator does well, what it can't do, and how to coax good results out of it.

Two song lengths

Clip, always 30 seconds. Cheap. Great for iterating on a prompt before committing.
Pro, typically a couple of minutes. You can nudge duration in the prompt but can't pin it precisely.

Safety filters. Prompts that name specific artists or request copyrighted lyrics get blocked. There's no workaround.
Watermark. Every generated audio file contains an inaudible identification watermark. It doesn't affect listening.
One-shot. You can't iteratively refine a clip with follow-up prompts. Each generation is independent, use "Remix this prompt" to restart from an existing recipe.
Variation. The same prompt produces a different song each time. To get something repeatable, be specific, name instruments, BPM, key, vocal style, and structure. The richer the prompt, the closer subsequent runs sound to each other.

Iterate with Clips. Cheap 30-second runs let you nail the prompt before spending Pro credits.
Be specific. Vague gives generic. Mention instruments, BPM, key, mood, and structure.
Use section tags. [Verse], [Chorus], [Bridge] give the model a structure to follow.
Separate lyrics from direction. Put lyrics in their own block, away from your musical instructions.

Describing the singer (gender, timbre, range) goes a long way. Some patterns that work well:

Female Soprano, clear, crystalline timbre, agile, soaring, airy high notes.
Female Alto, rich, warm, husky lower range, soulful, smoky with vocal fry.
Male Tenor, bright, piercing, energetic, youthful with a nasal edge, strong belting.
Male Baritone, deep, velvet-smooth, resonant chest voice, crooning delivery.
Weathered Rocker, raspy, gravelly, 90s-grunge texture, strained upper range.

Key / scale, "in G major", "D minor".
Mood adjectives, "nostalgic", "aggressive", "ethereal", "dreamy".
Duration cues (Pro only), "create a 2-minute song", or use timestamps in your lyrics.