Guide
AI Song Prompt Guide: How to Prompt for the Best Track
A good prompt is the difference between a generic AI song and a track that actually sounds like what you heard in your head. Here's how to write prompts for Toxic Sound that get the model to nail the vibe on the first try — and how to iterate when it doesn't.
The prompt formula
Every generation combines three inputs: genre, vocal type, and your theme prompt. The genre and vocal control the sound; the theme controls the words. Keep each one focused and they reinforce each other. Fight them and you get a muddled track.
Genre → instrumentation, tempo, rhythm Vocal → who sings and how Theme → what the song is actually about
Step 1 — Pick one genre, not five
One genre at a time. "Death metal country trap ballad" doesn't blend — it picks one and ignores the rest. If you want a hybrid feel, pick the dominant genre and describe the crossover in the theme (e.g. Metal + theme: "acoustic intro that explodes into a wall of distortion").
Step 2 — Match the vocal to the mood
- Male / Female — most tracks; picks a lead voice.
- Duet — call-and-response lyrics, breakup songs, storytelling.
- Choir — anthems, gospel, epic hooks, ritualistic energy.
Step 3 — Write a focused theme
The theme prompt is where most people go wrong. Long, vague, adjective-heavy prompts water the song down. Aim for one idea, one mood, one setting. Concrete beats abstract.
Weak
"A cool emotional song about life and love and pain and being real."
Strong
"3 a.m. drive home after a breakup, city lights, cigarette out the window, cold rage under the sadness."
Good themes usually include: who (narrator/character), where (setting), when (moment in time), and the emotional core. Two or three sentences is plenty.
Step 4 — Steer the delivery, not just the story
You can hint at delivery inside your theme: "shouted chorus," "spoken-word verse," "slow builder that drops at the last minute," "whispered hook." The model picks these up. Don't over-stack them — one delivery note per prompt keeps it clean.
Step 5 — Explicit content: say what you mean
Toxic Sound doesn't censor. If you want raw language, write the theme with the exact register you want — "furious, profane, no punches pulled" tells the lyricist to swing. Vague prompts return radio-safe results even here.
Step 6 — Iterate, don't overwrite
Same prompt run twice produces different takes. Before rewriting your prompt, hit Generate again. If two takes both miss the same way, then change one variable (genre, vocal, or theme) — never all three at once. That way you actually learn what moved the sound.
Prompt templates that work
Angry breakup anthem
Genre: Pop Punk · Vocal: Female · Theme: "Screaming out the sunroof after finding the texts. Fast, spiteful, chorus is a name you spit like a curse."
Late-night confession
Genre: Neo Soul · Vocal: Male · Theme: "Voicemail you'll never send. Warm Rhodes chords, hushed verses, one honest line in the bridge."
Ritual banger
Genre: Fantasy Metal · Vocal: Choir · Theme: "Army marching into a dying sun. Latin-sounding chants, war drums, hook the whole choir shouts together."
Common prompt mistakes
- Naming artists. The generator ignores band names; describe the sound instead ("gritty 90s grunge with dry drums").
- Wall-of-adjectives. Ten mood words average out to none. Pick two.
- Fighting the genre. "Death metal lullaby" will not resolve — pick one side and lean in.
- Novel-length prompts. Past ~2 short paragraphs the model starts skipping. Trim.
- Changing every knob at once. Move one variable per regeneration or you'll never know what helped.
Ready to try a better prompt?
Head back to the studio, pick one genre, one vocal, and write a two-sentence theme with a setting and a mood. Then generate twice before you change anything.