How do you make a song with AI, from the first line to a finished track?

1. idea         one sentence: who, what happened, one object
2. lyric        title + verses + chorus, with [Verse] / [Chorus] tags
3. description  genre, instruments, voice, mood, tempo word
4. generate     lyric in the lyrics box, description in the style box
5. pick         two to four takes, judged against a checklist
6. fix          rerun only the layer that failed

That is the whole route from nothing to a finished track, and it works the same way whichever AI music generator you end up using. Only step 4 happens inside the music tool. Steps 1 to 3 are writing, steps 5 and 6 are judgement, and most songs that come back sounding generic went wrong before anyone pressed a button: the idea was vague, the lyric was filler, or the description was a single genre word.

So the short answer to "how do you make a song with AI" is this: write down one specific idea, turn it into a tagged lyric, describe the sound in a short list of plain words, paste both into a generator that sings, compare a handful of takes, and repair whichever layer broke. The rest of this page walks each step with a real example, a brother moving to another city, carried from the first sentence to the fix.

Where we fit, before you pick a tool

We make one of those six layers. The lyrics generator at /generators takes a short form brief and returns a title and a full lyric with [Verse] and [Chorus] section tags already in place. It does not sing and it does not make audio, ever. The audio happens in a music generator you choose, usually Suno, made by a separate company: we do not resell Suno or speak for it. One generation costs 20 credits, which is $0.20; credits come only with a plan, and the plans are $5 a month (500 credits, 25 generations) or $19 paid once (1500 credits once, 75 generations, no monthly refill). The $5 plan is the cheapest way in.

If you already have finished words and only need the handoff into Suno, the narrower page is how to turn lyrics into a song. If you are still choosing the tool itself, which AI music generator suits which job sorts them by what you need at the end. This page is the whole trip, and it deliberately does not care which singer you hire at step 4.

Step 1: an idea is a sentence with a person and an object in it

"A song about my brother" is not an idea yet. It is a category. Hand a category to any generator, lyric or audio, and it will fill the gaps with the most average version of that category it can produce: brotherhood, always there for me, side by side. Nothing wrong, nothing yours.

An idea that survives the machine names three things: who the song is for or about, what happened that makes the song worth singing now, and one physical object that belongs to those two people and nobody else. Here is the brief this page will carry to the end:

My brother Dima moves to Vancouver in March. We shared a room
for twelve years and he always fell asleep with the radio on.
A send-off from me to him, hopeful, not sad.

That is 168 characters, well inside the 300-character theme field that step 2 uses. The object is the radio. The event is March and Vancouver. The person is Dima. Every later step leans on those details, and the song will never be more specific than this sentence is, because nothing downstream can invent your life for you.

A quick test before moving on: read the idea aloud and ask whether it could be about anyone else's brother. If yes, add the object. If you cannot find one, the song is not ready, and no amount of rerolling in step 4 will make it ready.

Step 2: the lyric decides what gets sung

Every generator that sings needs words. Some write their own if you leave the box empty, and those words tend to sound like everyone else's, because they were written from a genre and not from your sentence. Tools built mainly for instrumentals, a category that has included Soundraw and Mubert, are designed around the backing track rather than setting your words to a vocal line; features change, so check the current feature page of whatever you are using. Either way, the words are your job or your lyric tool's job, never the audio tool's strength.

You have two routes.

Write it yourself. Worth it if you already hear lines in your head. Keep verses to four to six short lines, give the chorus a single line that could be the title, and mark sections with tags in square brackets. Our guide to the song lyrics generator and where it breaks covers the second draft, which is where human judgement matters most.

Generate it from a brief. This is the step we sell. On /generators the idea from step 1 goes into the theme field (300 characters maximum), and the rest of the brief is picked from fixed options. For the Dima song it looks like this:

theme      My brother Dima moves to Vancouver in March. We shared a
           room for twelve years and he always fell asleep with the
           radio on. A send-off from me to him, hopeful, not sad.
language   en
style      Indie
POV        second
emotion    hope, longing
vocal      male
hook       title-line
age        any
structure  classic

Each field removes a guess. second means the song talks to Dima, not about him. title-line means the hook repeats the title, which gives the chorus something to come back to. Two emotions is the cap, and hope plus longing is the honest mix for a send-off. classic asks for verses, a chorus and a bridge, the shape most listeners expect from a three-minute song.

What comes back is a title and a tagged lyric. Here is a hand-written illustration of the shape to expect for this brief, the first verse and chorus only:

Leave the Radio On

[Verse 1]
Twelve years in the same small room
You'd fall asleep with the radio on
I learned to sleep through every song
I never learned to sleep without one

[Chorus]
Leave the radio on in Vancouver
Leave it playing low the way you do
March is coming and you're going
And I'll leave mine on for you

The room, the twelve years, the radio, March and Vancouver all come from the theme field. That is the lesson in one block: a lyric tool can only be as specific as the sentence you give it. Feed it "a song about my brother" and you get a song about a brother. Feed it the radio and you get a song that could only be about Dima.

Read the draft once with a pen before it goes anywhere. Cut any line you would be embarrassed to hear sung at the farewell dinner. Swap any word the singer might stumble over. If the name will be sung, make sure it is spelled the way it is pronounced, because a generator reads letters, not intentions.

Step 3: the description decides how it sounds

The lyric fixes the words. The description fixes the band, the singer and the weather of the track. Leave it empty and the tool falls back to its own average arrangement, which is how a good lyric ends up sounding like hold music.

Across tools the description is built from the same five ingredients:

• genre, one or two, not five: indie folk
• instruments, the two or three that carry the song: warm acoustic guitar, soft brushed drums
• voice, who sings and how: gentle male vocal
• mood, in words a listener would use: hopeful, bittersweet
• tempo word, not a number: mid-tempo

For the Dima song:

indie folk, warm acoustic guitar, soft brushed drums,
gentle male vocal, hopeful, bittersweet, mid-tempo

The shape of that text is where tools part ways. Suno's Style field reads a comma-separated list like the one above. Other generators ask for a sentence, offer only genre and mood presets, or split the description across several controls. Our prompt generator and our catalogue of nearly 2 000 styles were built for Suno's format, and we have not verified them in every other tool. When you move to a different generator, keep the five ingredients and adapt the form: if the tool wants prose, the same content reads as "a warm indie folk song with acoustic guitar and brushed drums, sung by a gentle male voice, hopeful and a little bittersweet, mid-tempo". The tool's own documentation is the authority on what it accepts.

Two habits that save regenerations whatever the tool:

1. Stop at seven or eight descriptors. Past that, the words start to contradict each other (aggressive next to soothing) and the result lands in the muddy middle of both.
2. Leave out the key and the exact BPM. Our Suno guides recommend tempo words over a number, and locking a key tends to fight the melody the model wants to write. If your tool's docs say otherwise, follow the docs.

Step 4: generate, with each text in its own box

This is the only step that belongs to the music generator, and it is mostly paste.

• The lyric, tags included, goes in the lyrics field. Everything outside square brackets is likely to be sung, so strip notes to yourself, chord names and stray comments before pasting. The guide to Suno metatags that work lists which tags Suno respects.
• The description goes in the style or prompt field.
• The title goes in the title field if there is one. It is also the chorus line, which helps the tool treat it as the hook.

If your tool has a single box and no lyrics field, check its documentation before pasting: some treat everything as a description and will write their own words over your idea. That is exactly the case where the lyric step was wasted, so it is worth two minutes of reading first.

Step 5: pick with a checklist, not a mood

Many tools return two or more takes per run. Resist the urge to press generate twenty times. Generate a small batch, then judge every take against the same five questions:

1. Can you understand the words on the first listen, without reading along?
2. Does the chorus arrive where the [Chorus] tag is, or does the song blur the sections into one long verse?
3. Is the name sung correctly, with the stress where Dima's family would put it?
4. Does the voice match the description, or did gentle male vocal come back as something else entirely?
5. Does the track end, or does it stop mid-line or run on into a repeat nobody asked for?

A take that passes all five is your song. A take that fails one is information: it tells you which layer to fix. Keep the best candidate even if it is not perfect, because the next run may be worse, not better.

Step 6: fix the layer that failed, not the whole song

The expensive mistake is to change everything at once, rewriting the lyric and the description and the tool, and then not knowing what helped. Every common failure lives in exactly one layer.

The words are hard to understand. Usually the lyric: lines too long for the phrase, too many syllables crammed into the melody. Shorten lines, cut filler words, regenerate.

The name comes out wrong. The lyric again. Spell it the way it sounds, or move it to a place in the line where it lands on a strong beat.

The chorus does not feel like a chorus. Check the tags first, then the lyric. A chorus that is as long and wordy as the verse gives the tool no reason to lift it. Cut it to four short lines built around the title.

The genre drifted. The description: probably too many descriptors or two that pull in opposite directions. Remove the weakest three and run again.

The voice is wrong. The description. Name the voice once, clearly, and delete any other word that implies a different singer.

The song runs too long or cuts off. Usually the lyric is simply longer than the track the tool makes by default. Cut a verse rather than hoping for a longer render. Some tools let you extend a track; for Suno the extend and edit guide covers it.

Everything is fine and it is still boring. That is step 1. The idea had no object in it, and the whole chain inherited the gap. Go back to the sentence, not to the button.

What "finished" means, and what it does not

A finished track is a file you can play at the farewell dinner. A released track is a different thing: it involves the generator's terms for your plan, a distributor if you want it on streaming services, and rights questions that depend on where you live.

On our side, the positions are published on the disclaimer page: commercial use of the lyric is allowed, we claim no authorship over what the generator writes, the rights you get are non-exclusive, and because one model writes for every user, similar briefs can produce overlapping lines. The audio tool has its own terms, and those vary by plan and change over time, so read the current page for the plan you are on before you publish anything. If you do go on to publish, putting Suno songs on YouTube and Spotify walks through the next stage.

This is a description of how these pieces usually fit together, not legal advice. For a release that matters, check the current terms of each service involved and, if the stakes are real, ask someone qualified in your country.

FAQ

How long does the whole process take? The writing steps are where the time goes. A clear idea takes a few minutes, a lyric draft seconds if generated and longer if written, and a description a couple of minutes. Generation itself runs in minutes on most tools. The fix loop is the variable part.

Can I make a song with AI for free? Many music generators have a free tier, and free tiers commonly limit how many songs you can make or what you may do with them commercially. The details change often, so read the tool's current pricing page. Our lyric generator is paid and has no free runs: each run costs $0.20 inside a plan, the cheapest plan is $5 a month (25 runs), and the other is $19 paid once, with 75 runs included once.

Do I need to write the lyrics myself? No. You can write them, generate them from a brief, or mix both: generate a draft, keep the lines that sound like you, rewrite the rest. What you should not do is leave the words to the music tool's defaults if the song is meant for a specific person.

Which generator should I use for the audio? It depends on the job, and the job decides more than any ranking does. This comparison by use case is the fastest way to choose. For a song with vocals and your own lyric, you need a tool with a lyrics field.

Can the song be in another language? Our lyric generator writes in nine: English, Russian, Spanish, German, French, Italian, Portuguese, Japanese and Korean. Whether the audio tool sings a given language well is a question for that tool; test with a short lyric before committing a whole song.

What if I only want an instrumental? Then skip step 2 and give the description everything. The lyric layer only matters when someone is going to sing.

Your next move

Take your idea and write it the way the Dima example was written: a person, an event, one object, under 300 characters. Then open the lyrics generator, paste that sentence into the theme field, pick a style, up to two emotions and a structure, and generate. Inside a plan each run costs $0.20 and returns a title and a tagged lyric ready for step 4. The cheapest way in is $5 a month, which covers 25 runs; the $19 one-time plan comes with 75 runs once, which is enough to rewrite the brief a few times and keep going until the chorus sounds like the two of you.