How to write an AI music prompt you can revise

Start with the song’s job

The examples are illustrative writing exercises. No recordings were generated or evaluated for this article.

“Emotional pop song” leaves a lot undecided. An emotional song for a birthday entrance and one for a late-night confession have different jobs.

Write the job first:

A song about a band returning to rehearsal after a rough first gig. The first verse is frustrated; by the last chorus, they are ready to try again.

That gives later decisions a purpose. You might choose a sparse opening to leave room for frustration, then add instruments as the band finds its confidence. It also gives you a listening question: does the song make that turn?

You can also use this approach for a song with no recipient. You can start with a scene, an argument, or a musical idea you want to explore.

Turn a mood into choices someone could hear

Use concrete descriptions where they matter:

These are possible interpretations, not fixed translations of emotion into sound. Your idea of “intimate” may be a piano and a voice; someone else’s may include quiet electronic textures.

Google DeepMind’s Lyria guide recommends describing musical elements including genre, tempo, instruments, and vocal qualities. It is guidance for Lyria; do not assume every app exposes all of the same inputs.

Describe a vocal without naming an artist

If the voice in your head belongs to a famous singer, describe the qualities you want to hear instead of asking for that person. Start with delivery, texture and phrasing. Add language, register or backing voices only when they matter to the song.

These are original writing examples, not tested formulas:

Google DeepMind’s Lyria guide describes vocal profiles through qualities such as register, texture and language. The examples here are our own. A description does not guarantee a matching voice, and a tool may not support each choice separately.

Give the important choice room to work. “Breathy, belted, whispered and full-power throughout” leaves the intended delivery unclear. If you want a change, locate it: a restrained verse followed by a stronger chorus. If a particular accent or language matters, have someone who knows it review the sung result rather than assuming the label did the job.

Build one coherent prompt

Here is the rehearsal idea with a musical direction:

An English-language indie-pop song about a band rehearsing after a rough first gig. The verse notices a broken guitar string and a set list covered in crossed-out songs. The chorus turns toward playing together again. Use a conversational vocal, warm electric piano, bass, and light drums. Begin with plenty of space; make the last chorus fuller and more assured.

The story and arrangement are aiming at the same change. There is no requirement to use this exact order or length. A short prompt may be enough when you already know what you want.

Before generating, look for accidental contradictions. “Keep it quiet throughout” and “finish with a huge explosive chorus” might describe a deliberate surprise, but they need a decision. If the contrast matters, explain where it happens. Otherwise, choose the direction that serves the song.

Separate lyrics from musical instructions

If a tool has separate fields for lyrics and style, put each kind of instruction in the field intended for it. If it accepts only a story or theme, do not assume pasted lyrics will be performed verbatim.

For a story prompt, make the point of view explicit:

The singer is a member of the band. Use “we” for the band, and keep that point of view throughout.

For a workflow that accepts your own lyrics, check its current formatting instructions. The Lyria guide, for example, documents a way to label supplied lyrics. For that Lyria workflow, an original example would be Lyrics: We left the last light burning / We came back for one more song. That syntax belongs to the documented workflow; it is not a promise that Gana or another app accepts verbatim lyrics in the same way.

Avoid packing private details into a prompt just to make it specific. “A broken string beside a crossed-out set list” gives a scene without adding a street address or employer’s name.

Copy a brief you can actually review

Replace the brackets. Keep the lyrics in the tool’s lyrics field when one exists; this is a planning worksheet, not special syntax every generator accepts.

Song’s job: [what the listener should understand or feel].
Story or theme: [one scene or idea].
Words: [original lyrics in their own field, or a theme for new lyrics].
Musical direction: [one main style, pace and a few instruments].
Voice: [delivery, texture, language and phrasing that matter].
Development: [what changes between the opening and ending].
Three checks: [the details you must hear before accepting it].

For the rehearsal example, the three checks could be: the story stays in the band’s “we” perspective; the words are clear in the opening; the last chorus feels fuller than the first verse. Write those down before listening. If a result misses one, record the miss before deciding what to change.

Exact duration, lyric delivery and arrangement still need checking in the actual recording. A request for a finish at a particular moment is a brief, not a verified edit. For a video or production project, use the delivery and purchase checklist to check file, ending and permission requirements too.

Keep a revision notebook

Save the original prompt before you change it. Then write what you heard and what you want to try next.

Use this note format after each attempt:

Record the tool, date, available model/version, settings, and cost or credits where relevant. Keep failed attempts in the notebook. They help you avoid repeating a change that did not solve the problem.

Changing one instruction at a time makes your intention easier to follow, but it does not make the result a controlled experiment. Generative systems can vary between attempts. A second recording may change in ways you did not request.

If you want to claim that a prompt change improves results, test repeated attempts using the same conditions and judge them against a criterion chosen in advance. One pleasing result can inspire your next song; it cannot establish a general rule.

Use the smallest revision that addresses the problem

If the words are hard to understand, try a less crowded arrangement or a clearer vocal delivery before rewriting the whole story. If the song feels emotionally flat, check whether the lyric and arrangement both stay in the same place throughout.

If the problem is a name, focus on pronunciation. If the problem is a generic message, add a scene. If the problem is the wrong sound, change the musical direction. Those are different failures and need different revisions.

Stop when the recording serves the song’s purpose well enough for its intended use. More generations cost time and may cost money; a later attempt is not automatically an improvement.

What this means in Gana

Gana’s current prelaunch flow starts with a written idea, a musical direction, and a private 30-second preview. That preview is an excerpt from the full recording; purchase unlocks the same recording. Older clip-only previews may use a separate generation flow. Listen to the whole unlocked song before deciding whether it serves your brief.

You can explore public examples on the songs page. For a gift, the personalized-song guide focuses on finding the scene worth writing about.

Coming soon for iPhone. Write your song’s job in one sentence, add the sound you want to explore, then see the planned iPhone workflow. Keep that first brief ready.