How to make a song with AI, step by step, with the prompts and lyrics that worked

To make a song with AI, decide who the song is for, pick how you start (a description, your own lyrics, a hum or a photo), write a prompt that names genre, tempo, instruments, voice and subject, add lyrics with section tags, compare two takes, then fix the weak part in a studio instead of starting again. The whole loop fits on a phone. Below is the method we use at Midhull, shown on songs we made with our own sound engine, with the exact prompts and lyrics next to the audio.

Updated

Midhull's Describe it screen turning a one line idea into a song plan with genre, tempo and voice

Midhull is a music studio app for iPhone and Android, launching soon in English, German, French, Spanish, Japanese and Korean. Everything on this page describes what is in the app today; where a feature has limits, we say so.

The short version

  1. Decide what the song is for and how long it needs to be.
  2. Choose your way in: describe it, write it, hum or sing it, snap a photo, dedicate it, tap a rhythm, remix or sample.
  3. Write a prompt with five parts: genre and era, tempo, instruments, voice, subject.
  4. Write or edit the lyrics with section tags and even line lengths.
  5. Listen to both takes and keep the one with the better chorus.
  6. Fix what is wrong with a targeted edit: one section, one line, the key, the tempo.
  7. Export the format your destination needs, at the right loudness.
A live band performing on stage under coloured lights

Step 1: decide what the song is for

A song for a friend's birthday and a song you want to release need different shapes, and the purpose should set the structure before you type a word.

PurposeLength that worksStructure we start fromWhy
Birthday or thank you message60 to 90 sVerse, chorus, chorusThe name lands in the chorus twice, and nobody waits a minute for it
Short video hook15 to 30 sChorus only, or pre chorus into chorusShort video audio is chosen in the first seconds
A song you want to keep and share2:30 to 3:30Verse, pre chorus, chorus, verse, pre chorus, chorus, bridge, chorusThe classic pop shape gives the chorus three chances and the bridge a turn
Instrumental for a video or a game60 to 120 sIntro, build, peak, outroMusic under picture needs a clear arc, not a hook

Midhull's longest song in our engine test ran 215 seconds (3:35) and sang every section in order: verse, pre chorus, chorus, verse, pre chorus, chorus, bridge, chorus. So a full length song is a realistic target from one set of lyrics.

Listen

Sunrise, take the long way

Summer pop · Bright

1:30

Leaving the small town

Alternative rock · Restless

2:30

Letter to an old friend

Acoustic ballad · Bittersweet

2:30

Porch light

Pop rock · Wistful

3:35

First light

Cinematic · Epic

1:30

Step 2: pick your way in

Most AI song generators give you one text box. Midhull has nine ways to start, because the best starting point depends on what you already have.

You haveModeWhat you giveBest for
A feeling or a sceneDescribe itA sentence or twoFirst songs, moods, instrumentals
Words you wroteWrite itYour lyrics, line by line, with lines you can lockPoets, songwriters, anyone with a message
A tune in your headHum itA few seconds of hummingMelodies you can hum but cannot write down
A vocal ideaSing itA sung phraseSingers sketching an idea
A pictureSnap itA photoTrips, pets, a night out
A personDedicate itAnswers to a few questions about themBirthdays, anniversaries, thank yous
A grooveTap itA rhythm tapped on the screenBeat first writers, rap
A song you love on MidhullRemix itA community song, credited automaticallyNew versions, other genres
A sound you recordedSample itA chop of your own in app recording, a community song or a Midhull packBeatmakers and rappers

Sample it only takes your own in app recordings, community songs whose owners allow sampling, and Midhull's own sample packs. Commercial recordings cannot be uploaded to sample. That rule keeps your song yours.

In the app

The Create screen in Midhull with nine ways to start a song
Writing lyrics in Midhull with section tags and a locked line

Step 3: write the prompt

A good prompt is a production brief. Five parts carry nearly all of the result:

PartWeakStrong
Genre and erapop2010s summer pop
Tempoupbeat118 BPM, or "mid tempo" if you do not know numbers
Instrumentsnice musicbright synths, handclaps, clean electric guitar
Voicea singerfemale lead, close and breathy in the verses, full in the chorus
Subjecta happy songdriving to the coast with friends at sunrise

Here is the prompt behind one of our test songs, "Sunrise, take the long way":

Upbeat summer pop, bright synths and handclaps, female lead, about driving to the coast with friends at sunrise.

And the one behind "Leaving the small town", made from a description only:

Anthemic alternative rock song, big distorted guitars, driving drums, raw male vocal, about leaving a small town behind.

Three habits that improve every prompt:

  • Name instruments, not adjectives. "Fingerpicked guitar and soft piano" tells the engine more than "beautiful and emotional".
  • Give one subject, with one concrete detail. "A letter to an old friend" beats "friendship"; "coffee's cold but the night is gone" beats "late night".
  • Steer by genre, era, mood and instruments, never by an artist's name. Midhull does not imitate real artists, so prompts that name one are steered back to style words. "Late 70s soft rock, warm Rhodes, close harmonies" gets you there without borrowing anyone's identity.

Step 4: write lyrics that can be sung

Lyrics on a page and lyrics in a song are different animals. The engine has to fit your words to a melody, and it does that best when the lines give it a pattern to follow.

Use section tags. Put each section on its own block with a tag in square brackets: [Verse], [Pre-Chorus], [Chorus], [Bridge], [Outro]. Midhull understands intro, verse, pre chorus, chorus, post chorus, bridge, rap, drop, breakdown, solo, interlude and outro.

Keep line lengths even inside a section. Four lines of 7 to 9 syllables in a verse give a steady melody. A 15 syllable line in the middle of short ones gets rushed.

Rhyme at the ends of lines 2 and 4 at least. It helps the melody resolve.

Repeat the chorus word for word. Repetition is what makes a chorus a chorus.

These are the lyrics we sent for "Sunrise, take the long way":

[Verse]
Windows down and the radio on
Coffee's cold but the night is gone
Five of us in a borrowed car
Chasing light to wherever we are

[Chorus]
Sunrise, take the long way home
Sunrise, we were never alone
Hold the moment, let it glow
Sunrise, sunrise, don't let go

Every verse line is 7 to 8 syllables, the rhymes sit at the line ends (on, gone; car, are), and the chorus opens three of its four lines on the same word. We then ran the result through Whisper large v3, a speech recognition model, to check how many words came out right: the word error rate was 0.04, which means about 4 words in 100 differed from what we wrote, and the only clear slip was "let it glow" heard as "let it go".

If you do not want to write lyrics at all, Describe it writes them for you with Midhull's own lyric writer, in the song's language. You can then edit any line and lock the lines you like.

Step 5: compare two takes

Every song you make comes back as two takes to choose from. Listen to both choruses before you listen to anything else: the chorus decides whether the song works. Previews do not use your balance; a song counts only when you keep it, and if a take comes out broken (a gap in the vocal, a skipped line, clipping, a song cut short, the wrong language or silence) the song goes back to your balance automatically.

Step 6: fix the part, keep the song

This is where most AI music workflows fall apart. You like the song, but the second verse is flat, and the only tool is "generate again", which throws away the parts you liked. Midhull's studio lets you change one thing at a time:

What is wrongWhat to do in the studio
The chorus is flat but the verses are greatRegenerate that section only; the rest stays exactly as it is
One word is sung badlyRe sing that one line
You want different words in one lineRewrite that line; the rest of the song stays
It ends too soonExtend the song, 10 to 90 seconds at a time
It needs a turn before the last chorusAdd a bridge
It sits too high or too low for the moodChange the key, up to 6 semitones either way
It drags or rushesChange the tempo, anywhere from 40 to 220 BPM
Wrong voice for the songChange the vocal style: female, male, duet, group, choir, rap, whisper or your own verified voice
Too producedMake it acoustic
Right song, wrong genreRemix it into another genre and keep the vocals

Songs also stay on the engine version they were made with, so an edit next month still sounds like the song you made today.

Studio tools open up as you make songs and level up: rewriting and re singing a line are there from your first song, key and tempo open at level 3, regenerating a section and adding a bridge at level 5, swapping an instrument at level 8 and stem export at level 10. Extending a song, changing the vocal style, making it acoustic and remixing into another genre are open from the start.

Step 7: export for where it is going

DestinationFormatLoudness preset
Streaming servicesWAV-14 LUFS
Short videoMP3-11 LUFS
YouTubeWAV or MP3-14 LUFS
Podcast introMP3-16 LUFS
A producer or a DAWStems (drums, bass, keys, lead; stem export opens at level 10) plus a chord MIDI file, on plans that include themOriginal
A singer who wants to learn itLyric sheetNot applicable

LUFS is the standard measure of perceived loudness. Matching the target means a platform does not turn your song down or squash it. Downloads of songs you make are never locked later, even if you change plans.

Three complete examples

A ballad from a description. Prompt: "Acoustic ballad, fingerpicked guitar and soft piano, tender female vocal, about a letter to an old friend." Result: a 2:30 song with verse, chorus and an acoustic guitar bridge. Listen to "Letter to an old friend".

A full length song from your own lyrics. Prompt: "Heartfelt pop rock, acoustic and electric guitars, piano, full band, warm male vocal, big singalong chorus, emotional bridge." Lyrics: our own, eight sections. Result: "Porch light", 3:35, every section sung in order, word error rate 0.06. The bridge is the moment to listen for:

[Bridge]
Maybe we were young and maybe we were proud
Maybe all we needed was to say it out loud
I am saying it now, I am saying it now

An instrumental. Prompt: "Cinematic orchestral build, soaring strings, French horns, taiko drums, epic and hopeful, no vocals." Result: "First light", 90 seconds with no vocal at all (the vocal stem measured 34 dB under the mix, and speech recognition found no words).

Common mistakes and the fix

MistakeWhat happensFix
Ten adjectives, no instrumentsA generic mixName two or three instruments
A whole story in one verseRushed, crowded linesOne idea per section, 4 lines each
No chorus tagThe song wandersTag it and repeat it word for word
Mixing two genres without a leadNeither genre comes throughLead with one, add the other as colour: "lo fi hip hop with a bossa nova guitar"
Regenerating a whole song to fix one lineYou lose the parts you likedRe sing or rewrite that line
Sunrise, take the long wayLeaving the small townLetter to an old friendPorch lightFirst light

Frequently asked questions

Can I make a song with AI on my phone?

Yes. Midhull is built for iPhone and Android. You can describe a song, write lyrics, hum, take a photo or tap a rhythm on the phone itself, then edit and export there.

Do I need to know music theory?

No. Tempo in BPM and the key are optional. "Mid tempo" and "in a sad key" are fine starting points, and you can change key and tempo after you hear the song.

Can I use my own voice?

Yes, once you verify it. Midhull's own voice feature takes a live read of a consent phrase, checks it, and then lets the song be sung in your voice. It only works with your own voice.

What languages can the songs be in?

We have tested singing in 37 languages, including Spanish, Portuguese, Hindi, Arabic and Filipino. The app itself is in English, German, French, Spanish, Japanese and Korean. See AI songs in other languages for the full list and the measured results.

Is the first song free?

Yes. You make your first full song while you set up the app, before any payment. After that, songs come with a plan. Prices are published at launch.

Who owns the song I make?

Every song carries a rights card that says, in plain words, what you can do with it on your plan. For the wider question of AI music and copyright, read Can you use AI music commercially?.

Where to go next

Midhull is almost here

We’re finishing the app for iPhone and Android. Join the waitlist and we’ll tell you the day it opens.

Coming soonApp Store Coming soonGoogle Play
Which phone do you use?

We use your email only for the waitlist and launch news. Unsubscribe from any email in one tap. See our privacy policy.