Most weak AI songs come from weak briefs, not weak models. "A sad song about my ex" gives the model a hundred directions and it picks the average one. This guide is the set of instructions we give ourselves before we tap Create in Hitz AI. It works in any generator, and the examples use the fields you will find in the app.
Start with the shape, not the story
Before the subject, decide three things: how it should feel, how fast it moves, and who is singing. Those three set the whole track.
- Feel: warm, cold, bright, heavy, airy, tense, tender.
- Pace: slow and spacious, walking pace, driving, frantic.
- Voice: a soft female lead, a rough male lead, a duet, no vocals at all.
In Hitz AI, the mood picker covers the feel with 19 options, and the vocal picker offers female, male or duet. Instrumental only turns the vocal off.
Name the genre, then name the neighbourhood
Genre alone is a wide net. "Pop" covers stadium pop and bedroom pop recorded on a laptop. Add one qualifier that says where in the genre you live.
- "Indie folk, close-mic acoustic guitar, a little tape hiss."
- "Afrobeats, log drum, sunny, made for a rooftop."
- "UK drill, sparse, cold synth, half-time hats."
- "Country, modern Nashville, big chorus, pedal steel underneath."
In Hitz AI, pick one of the 21 preset genres, or choose Custom and write the neighbourhood yourself. On V8 Fast, V8 Pro and V8 Evo the custom sound can run to about a thousand characters; on V8 Raw keep it under two hundred.
Describe the instruments like a producer would
Say what plays, and what it does. "Piano" is fine. "Felt piano playing sparse chords, nothing on the beat" is a direction. Three or four instruments is plenty. Past that, the model starts averaging again.
Say what is missing too. "No drums until the second chorus" and "no synths" are among the most useful instructions you can give, because the default is to fill every gap.
Give the vocal a character, not just a gender
"Female vocal" is a checkbox. A character gives the model a performance to aim for.
- "Breathy, close to the mic, almost spoken in the verses."
- "Confident, belted chorus, gospel-style harmonies behind it."
- "Deadpan, low, a little bored, like he has said this before."
If the song is for someone, say how the singer feels about them. That one line does more for the delivery than any adjective about tone.
Lyrics: give the model a scene, not a summary
When Hitz AI writes the lyrics for you, hand it details it could not invent. A summary ("we broke up and I miss her") produces generic lines. A scene produces lines only you could have written.
Instead of: "A song about missing my best friend who moved away."
Try: "My best friend Sam moved to Lisbon in March. We used to get bad coffee at the gas station on Route 9 every Sunday. I still drive past it. Keep it funny, and mention the coffee."
Names, places, objects and one rule about tone. If you already have lyrics, choose My Own Lyrics, paste them in and spend your brief on the sound instead. See how to write lyrics with AI for more.
Ask for a structure when it matters
Most songs do not need one. When you have a picture of the arc, say it in one line: "Quiet verse, the chorus opens up, a short bridge, then the last chorus with everything in." For instrumentals, say where the energy should peak and how it should end. "Ends on a single held note" beats a fade-out you did not ask for.
What to leave out
- Artist names. Hitz AI does not imitate real artists, so a name gets ignored or steers you somewhere odd. Describe the sound instead: "early 2000s R&B, live bass, Rhodes chords."
- Contradictions. "Chill but high energy" hands the model a coin to flip.
- Everything. A brief with twelve instruments, four moods and three genres is not a brief. Pick the three things that matter and let the model handle the rest.
Pick the model for the job
Hitz AI has four models, and you pick one per song:
- V8 Fast: a quicker take that still keeps the musical detail. Two takes per song.
- V8 Pro: lifelike vocals and rich arrangements, with the most control over your sound. Two takes per song.
- V8 Evo: wilder, less predictable readings of your lyrics and style. Two takes per song.
- V8 Raw: simpler arrangements, and the least likely to be held up by content limits. One take per song.
Two takes are the fastest way to hear what your prompt is actually controlling: whatever stays the same across both came from your words. More in the four models.
What you can write about
A lot of generators refuse more than they need to, and a breakup song that is not allowed to be bitter is not much of a breakup song. Copyrighted lyrics, impersonating real people, sexually explicit content, harassment and hate stay off limits, as the terms set out. Inside those lines, Hitz AI tries not to second-guess your song. A diss track for the group chat can land. An apology can be uncomfortable. A birthday song can be a little unhinged. If a prompt still gets held up, try V8 Raw.
Three briefs you can paste
A birthday song for a friend
Bright pop punk, fast, gang vocals in the chorus, one guitar and live drums. Lyrics: it is Maya's 30th, she is scared of getting old, she still drives the same Honda from college, and she once got kicked out of a karaoke bar in Austin. Loving, not mean.
Something to drive to at night
Slow synthwave, wide analog pads, a simple arpeggio, instrumental. Builds once around the middle, then thins out and ends on the pad alone.
A quiet apology
Indie folk, one acoustic guitar, close-mic soft male vocal, a little room noise. Lyrics: I forgot her mother's memorial, I am not asking for forgiveness, just telling her I know. Plain words, no big metaphors.
Get Hitz AI on the App StoreFree to start. Your first songs are on us.
Questions
How long should an AI song prompt be?
Two to four sentences. Enough to name the feel, the pace, the voice and two or three instruments. Longer briefs work only if every line adds a decision.
Can I use an artist's name in the prompt?
It rarely helps. Hitz AI does not imitate real artists. Describe the sound you associate with them instead.
Why does the same prompt give different songs?
Generation is not deterministic, so treat each run as a take. In Hitz AI, three of the four models return two takes, which makes it easy to hear what your prompt controls.