AI songwriting: a bad first writer, a very good second

I asked a model for twenty opening lines on the same brief, three separate times. Sixty lines came back. Fifty-one were a variation on "city lights blur past the window". Six were competent and dead. Three had something in them, and one of those three contained the word thermostat, which no chorus has any business containing. That line opens a finished song now.

Three usable out of sixty is a humiliating hit rate for a writer and a perfectly good one for a machine, because the sixty took four minutes. Writing three openings by hand takes an afternoon, and two of them will still be the city lights.

So here is the position this piece defends the whole way down: a language model is a bad first writer and a very good second one. It has no taste, and it has no memory of what actually happened to you. Those two absences are exactly why it cannot choose the true line. They are not why it cannot produce forty candidates, and choosing from forty is a skill you already have.

Everything below is about the five places in the process where that trade pays, and the one place where it quietly costs you the song.

The blank page: ask for twenty firsts, not for a song

Never ask a model to write the song. Ask it for twenty first lines and nothing else.

The prompt that works is mostly a list of bans. Unconstrained, the model reverts to the mean, and the mean of every song ever written is weather, roads, and shadows. So: twenty opening lines, present tense, no metaphor involving weather, night, driving or mirrors, no line beginning with a gerund, each line a complete grammatical sentence. Then read them fast.

You keep two or three. That is the correct number. On a good run I keep four, and I have thrown away all twenty twice, which cost me four minutes rather than a morning.

What to ignore: any line that scans perfectly on the page. Perfect scansion is the model showing off, and a line that reads like verse usually sits badly against a melody, because melodies want a stumble somewhere. Also ignore anything that already contains the emotion word. "I'm lonely in this house" is not a first line, it is a summary. If you want the deeper version of that distinction, how to write song lyrics is the piece that pulls it apart properly.

The second verse that always sags

Verse one carries the situation. The chorus makes the claim. Then verse two arrives with nothing left to say, so it restates verse one with different nouns and everybody can hear it.

The instinct is to ask the model for a second verse. Do not. You will get verse one paraphrased, because the average of every second verse ever written is verse one paraphrased.

Ask for the job instead. Paste verse one and the chorus, then: "list six things verse two could do that verse one has not already done. Do not write any lines." What comes back is roughly this every time, and it is useful every time: move time forward, change who is speaking, bring in the second person, name the cost, contradict the chorus, pull the camera back. Six options, one of which you had not considered at 1am.

Then you write the verse. The model picked the angle, you wrote the words. That split is the whole method.

What to ignore: any actual lyric lines it volunteers when you asked for angles. It will volunteer them. They are always the paraphrase.

The line you have rewritten nine times

There is one line in most songs that will not sit. I had "and I never learned to stay" through nine versions across two weeks, and the ninth was worse than the first.

The problem is almost never the meaning. It is the syllable count fighting the melody, or the stress landing on the wrong beat. That is measurable, which makes it the one thing the model is genuinely good at.

Give it the shape, not the vibe: "rewrite this line twenty times. Seven syllables exactly. Stress on syllables one, three and five. End on a hard consonant. Same meaning. Do not comment." Twenty, not five. Five gives you the five most probable rewrites, and probable is the enemy here.

What to ignore: anything the model offers to "improve" about your imagery, and any version where it has quietly added a syllable to make its own line prettier. Count them yourself. Models are mediocre poets and competent counters, and you should hire them for the counting.

The title you cannot find

The title is usually already inside the lyric, and you cannot see it because you have read the lyric two hundred times.

Do not ask for title ideas. Ask a model to extract, from your finished lyric, every noun phrase of two to four words, then rank them by how rarely a phrase like that appears in song titles. Rarity is the entire selection criterion. Ask for titles in the open and you get "Echoes of Yesterday", which is the average of all titles and therefore the worst possible one.

That extraction move works because it constrains the model to words you already wrote. Nothing new gets invented, so nothing generic gets in. If you want the wider set of places real titles come from, song title generators covers the ones that are not extraction.

The second language version

A translated lyric is usually semantically correct and rhythmically dead. The meaning survives, the melody does not, and the vocal comes out mumbling because the stresses have moved.

Ask for a singable version rather than a translation, and say so in those words: same number of syllables per line, stresses in the same positions, and meaning is allowed to drift. Say the drift is allowed out loud, otherwise the model protects the meaning and wrecks the rhythm, every time. Then sing it yourself against the melody before it goes anywhere near a generator.

Suno mispronounces exactly where the written stress fights the melodic stress, which is a diction problem you can hear before you spend a credit. Making songs in other languages goes into what survives the crossing and what does not.

What you actually lose

Here is the counter-argument, and it is a real one, not a strawman I set up to knock over.

Every generated line is the average of everything written on that subject. Average is precisely what a lyric must not be. The whole value of a song is that one person noticed one thing that the rest of us walked past, and a system trained to predict the next most likely word is structurally incapable of noticing. Feed it enough of your draft and it will sand your song down into the shape of every other song about that feeling. You will not notice, because sanded-down feels good. It feels finished.

It averages structure too, not just lines. Left alone it produces four-line stanzas, ABAB rhyme, and a chorus that says the title three times, because that is the centre of the distribution.

Two defences, and I hold both of them hard.

Never keep a generated line you cannot say out loud in your own voice. Not sing. Say. Out loud, in the room, in your own accent. If it sounds like it came out of someone else's mouth, it did, and the listener will hear the seam even if they cannot name it.

Always replace at least one image per verse with something only you would know. The thermostat. The number of the bus. The brand of cigarettes your father smoked. The model cannot generate these because it was never in the room, and one such detail per verse is enough to make the whole thing yours. This is also the cheapest craft upgrade available: it costs nothing and it is the difference between a song about heartbreak and a song about your heartbreak.

The rest of the craft, the rhyme discipline and the chorus mechanics, is work you still have to do. No tool covers it.

The workflow that runs end to end on this site

Be clear about what these tools are. Both of them write text. Neither one produces audio, ever. The audio is Suno's job, and the words are ours.

Step one, the deterministic part. The Suno prompt generator is not a model. You tick slots and it returns the same recipe every time for the same slots: a Style line, a lyrics scaffold, a vocal block, and the Exclude tags. That last part matters more than it sounds. Suno reads a negative instruction as a suggestion, so writing "no autotune" in the Style field is a coin flip. The generator converts every "don't" into a positive tag sitting in Exclude, where it is actually enforced. Determinism is the point: when a track comes out wrong you can change one slot and know the change came from that slot.

Step two, the words. The AI song generator writes a title and a full lyric from a brief. The brief is deliberately narrow. One free-text field, theme, capped at 300 characters, and then choices: point of view (first, second, third, collective, narrator), song language from nine, target audience as a run of up to three adjacent age bands, structure (classic, no-bridge, double-chorus, story, loop, ballad), hook direction (title-line, question, contrast, confession, chant), vocal (unspecified, female or male), one of 53 root music styles, and one to three emotional tones.

There is no duet option, and that is a decision rather than a gap. Suno rarely splits the lines where you asked it to, so a duet brief tends to come back as one voice reading both parts.

The audience field takes a run, not a scatter. "18 to 34" works. "Kids and pensioners" is not representable, because a lyric written for both is a lyric written for neither.

A worked brief, exactly as I would fill it: theme last shift at a job I already quit in my head, first person, English, audience 25 to 44, structure story, hook direction confession, male vocal, root style Americana, tones longing and defiance. What comes back is a title plus a complete lyric already split into [Verse], [Chorus] and [Bridge], which is the format Suno's Lyrics field expects. That run costs 20 credits, and at 100 credits per dollar that is $0.20 per song, though check current pricing rather than trusting a number in an article.

Step three. Paste the lyric into Suno's Lyrics field and the Style line into Style. Then run the two defences on the lyric before you generate, not after, because a bad line does not improve by being sung. If you are unsure which tags Suno actually respects inside the lyric box, the metatag list is the reference.

Can you sell what comes out

Honestly: it depends on your tool's terms and on how much of the work is yours, and anyone giving you a flat yes or no is selling something.

Two practical things instead of a legal conclusion. Read the terms of every tool in the chain, separately, because the lyric tool, the music tool and the distributor each have their own, and they disagree more often than you would like. And keep your drafts, dated, with the human edits visible. If the question ever gets asked seriously, a folder showing nine versions of a line in your own handwriting or your own commit history answers it better than an argument does. Selling Suno songs walks through where the actual line sits.

FAQ

Will a listener know the lyric was AI-assisted? Not from the assistance. From the averaging. Listeners cannot detect a model, but they detect genericness instantly, and they call it "sounds like everything else". If you kept the specific details in, nobody asks the question.

Should I generate the whole lyric or write it myself? Generate the whole thing when you have no draft and need momentum, then treat it as raw material and expect to rewrite half. Write it yourself when the song is about something that happened to you, and use the model only on the second verse and the stuck line. The second mode produces better songs. The first mode produces songs at all, which some weeks matters more.

How many alternatives should I ask for? Twenty for lines, six for angles, one for structure. The number matters because the first three outputs are always the most probable ones, which is another way of saying the most generic ones. Interesting material lives around position eight.

Is a songwriting app worth paying for if I already have a chatbot? A chatbot will write you a lyric. It will not hand you a lyric already tagged with [Verse] and [Chorus], in the structure you asked for, with the hook aimed where you said, and it will not produce the same Style line twice for the same request. If you like tinkering with prompts, the chatbot is enough. If you want the same result on Tuesday that you got on Monday, determinism is what you are actually buying.

Does the AI song generator make music? No. It writes a title and a lyric. Suno makes the audio. Nothing on this site outputs a sound file.

What if every generated line is unusable? Your theme field is doing too little work. "Love" produces averages because it is a category, not a situation. "The last text I did not send" produces something, because it has a time, an object and a decision in it. 300 characters is a lot of room, and most people use nine of them.

Can I use it for rap, where the constraint is flow rather than melody? Yes, and the syllable-shaping trick from the stuck-line section matters more there, not less. Set the count and the stress positions explicitly and ignore the model's rhyme suggestions, which lean on the same four rhymes every rapper already retired.