When you ask a person to make a song, you can hum, point at a reference and talk it through over chai. An AI music tool only gets what you type. That typed description is the prompt, and it is the single biggest thing you control.
This guide explains what a prompt actually does inside an AI music generator, what to put in one, how it differs from your lyrics, and the common mistakes that turn a good idea into a confusing result. You do not need any music training to follow it.
What a prompt actually does
An AI music model has learned, from a very large amount of music paired with descriptions, how words like 'lofi', 'dhol' or 'melancholic' tend to sound. When you write a prompt, the model uses those words as conditions: it generates audio that is likely to match them. That is why the words you choose matter so much.
A prompt steers; it does not dictate. You cannot write out a melody note by note in a prompt, and the same prompt will give you a slightly different song every time, because the model draws a fresh take on each run. Think of it as briefing a talented session band that has never met you. A clear brief gets you close; a vague one gets you their best guess.
What to put in a good prompt
Most good prompts cover the same handful of things. You do not need all of them every time, but the more of these you pin down, the less the model has to guess.
- Genre or style: Bollywood romantic, Punjabi bhangra, Marathi lavani, Tamil kuthu, ghazal, lofi, EDM, devotional bhajan.
- Mood: joyful, nostalgic, heartbroken, festive, calm, triumphant.
- Tempo or energy: slow, mid-tempo, fast, or a number such as 90 bpm if you know it.
- Key instruments: two to four is plenty, for example dhol and tumbi, or harmonium and tabla.
- Voice: male or female vocal, soft or powerful, solo or with a chorus, or no vocals at all.
- Language: the language the song should be sung in.
- Use: wedding entry, birthday surprise, Instagram reel, morning aarti. The purpose quietly shapes everything else.
Prompt, style tags and lyrics are different things
Beginners often pour everything into one box. It helps to separate the jobs. The description or style prompt says how the song should sound. The lyrics say what is sung. Mixing the two can make the model sing your instructions out loud.
On Autunes you can describe the song in a sentence, add short comma-separated style tags such as 'acoustic, warm, 80 bpm', choose the language, and either paste your own lyrics or have lyrics written from your idea. Section tags inside the lyrics, like [Verse], [Chorus] and [Bridge], shape the structure of the song. If you only want music for a video, switch on instrumental only and the prompt does all the work.
Weak prompts and better ones
The fastest way to understand prompting is to see a vague idea turned into a usable brief.
- Weak: 'Punjabi song'. Better: 'High-energy Punjabi bhangra for a sangeet, dhol and tumbi, male vocal with group chants on the hook'.
- Weak: 'sad song'. Better: 'Slow Hindi heartbreak ballad, solo piano and soft strings, gentle female vocal, rain-at-night mood'.
- Weak: 'bhajan'. Better: 'Peaceful Krishna bhajan, harmonium, tabla and manjira, calm male vocal, call-and-response chorus'.
- Weak: 'background music'. Better: 'Upbeat instrumental for a food reel, ukulele and claps, bright and playful, no vocals'.
- Weak: 'Tamil song'. Better: 'Fast Tamil kuthu dance track, thappu drums and nadaswaram-style lead, festive group vocals'.
Common prompt mistakes
A few habits waste more credits than anything else. They are easy to fix once you know them.
- Contradictions: 'calm aggressive EDM lullaby' forces the model to pick, and you will not like its pick.
- Instrument overload: listing ten instruments usually gives a muddy mix. Choose the few that define the sound.
- Naming real singers: describe the qualities you like instead ('husky, soulful male voice'). Many tools ignore or block artist names, and imitating a specific person's voice raises legal and ethical problems.
- Lyrics in the style box: keep words to be sung in the lyrics field, or they may be ignored or sung strangely.
- Changing everything at once: if a result is close, change one thing in the prompt and try again, so you learn what made the difference.
What a prompt cannot do
Be realistic about control. A prompt cannot guarantee an exact melody, an exact song length to the second, or that every word is pronounced perfectly. Pronunciation is the area where AI vocals still slip most, especially with rare words, names and mixed-language lines.
That is why it helps when a tool checks the result for you. Autunes transcribes every finished vocal and compares it with your lyrics to flag lines that came out wrong. On Pro and Creator you can re-sing just that line with an edit section (3 credits) or run auto-fix pronunciation (10 credits), instead of throwing the whole song away.
How to iterate without burning credits
On Autunes a song costs 10 credits on Zori5 Turbo or 15 on the richer Zori5.5 XL, and the free plan gives 15 credits once on signup, which is enough for one song. Free users get XL only when they paste their own lyrics. So it pays to think before you press generate.
Write the prompt, read it back as if you were the band, and remove anything vague. If the first result is close but not right, a variation (5 credits) gives a fresh take of the same song; the free plan gets one, Pro and Creator get unlimited. Keep a note of prompts that worked. After a few songs you will have your own small library of briefs for weddings, reels and festivals that you can reuse and tweak.