How to write AI music prompts

Most AI song tools give you some version of three inputs: a free-text description of the song, a short style line that sets the sound, and an optional lyrics box. Mixing them up is the most common reason a prompt misfires. A description can tell a story and set a scene. A style line works best as a compact list of musical terms. The lyrics box should hold only words meant to be sung, plus section tags.

In the AI Song Generator Studio, Simple mode starts from your description and writes the lyrics for you, while Custom mode takes your exact lyrics and leans on the style line for the sound. Both modes also have a style field and an optional Exclude field, which is where most of the advice below applies.

A prompt that covers these five things leaves the model far less to guess. Order matters less than coverage, but leading with the genre helps, because the genre shapes every other choice.

When you supply your own lyrics, section tags tell the model where the energy should change. Put each tag in square brackets on its own line, directly above the lines it covers. Generators read tags as strong hints rather than strict commands, so a tag cannot guarantee a guitar solo, but it usually shapes the melody and the arrangement of that section.

Keep production notes out of the lyrics box. A line such as 'make this part louder' may simply be sung. Move notes like that to the style line, where the model expects them.

Negative words inside a positive prompt often backfire. Writing 'no drums' can still pull drums into the track, because the word drums is now in the prompt. Where a tool has a separate exclusion field, use it. In the AI Song Generator Studio that field is called Exclude, and it takes a short list such as 'heavy distortion, spoken word, auto-tune'.

Leave out real artist and band names as well. Describe the era, the instruments, the vocal texture and the production instead. Prompts that name real artists can be rejected by moderation, and imitating a specific person raises rights questions you do not want attached to a release.

Finally, drop instructions the model cannot act on, like 'make it go viral' or 'radio quality'. They use up space without changing the sound.

For a style line, somewhere between eight and twenty words is a comfortable range: enough room for the five ingredients, short enough that nothing contradicts anything else. A description can run longer when you are telling a story, but put the musical facts in its first sentence. If a long prompt gives muddled results, cut it in half and listen for which half the model was actually following.

Each generation is one draw from the space your prompt describes, so judge a prompt by a few results rather than one. Keep a short log of the prompt, what you liked and what was wrong, then change a single ingredient. If the groove is right but the singer is wrong, edit only the vocal words. If the chorus never lifts, add 'big chorus, stacked harmonies' instead of rewriting the whole line.

Once a take is close, stop prompting and start editing. Extend continues a track from a point you choose, replace section redoes one span, and cover or remix moves a finished song into another style. Every action in the Studio shows its credit cost before you confirm, so you can weigh an edit against a fresh generation.

A general chat assistant is a handy partner for prompt ideas, especially when you know the feeling you want but not the vocabulary for it. Ask it for ten style lines for a mood, ask it to translate a scene into instruments, or paste a weak prompt and ask which of the five ingredients is missing.

A request that tends to work: 'Give me five short style lines for a rainy late-night song, each with a genre, a BPM, three instruments and a vocal description, and no artist names.' Treat the answers as raw material. Assistants like to stack adjectives and sometimes invent genre labels no music model recognizes, so trim each suggestion back to concrete terms and remove any artist names that slip through.

Video generators read prompts differently from music generators. Instead of genre and tempo, they need a visible scene: who or what is on screen, where it is, what moves, how the camera behaves and what the light looks like. Write one short prompt per shot and line the cuts up with the song's sections, so the verse, chorus and bridge each get their own look.

The AI Song Generator Studio currently creates visualizer videos for tracks in your library rather than scene-by-scene videos, so take shot prompts like these to whichever video tool you use, with the finished song as your timing reference.