← Back to blog

How to write a song with ChatGPT or Gemini: prompts to copy

Published

A full songwriting session in ChatGPT or Gemini: set up the chat, write a song card, build the chorus and verses with copyable prompts, then make the melody yourself.

A songwriter at a desk crosses out and rewrites lines in a paper notebook next to an open laptop showing only soft abstract shapes, with an electric guitar on a stand behind, lit by window light and a desk lamp

To write a song with ChatGPT without it sounding generic, don’t type “write me a song about my sister”. Set up one chat or project for the song, paste a short song card (what happens, who sings to whom, two real details, style, syllables per line, banned words), then work one small job at a time: ten titles, eight options for one weak line, a rhyme list, an outline for a verse. Write the lines yourself from what you keep, check the syllables, and make the melody by humming over a chord loop and recording it. The same prompts work in Gemini and Claude. ChatGPT gives you text; if you want to hear a sung track, Suno or the music feature in the Gemini app generate one, with trade-offs you should know first.

This page is a hands-on session from an empty chat to a recorded chorus, with every prompt ready to copy and a worked example. If you want more prompt ideas for single lines, read how to write lyrics with AI without sounding generic; for the full route with music generators, see how to make a song with AI and keep it yours.

StepWhat you typeWhat stays yours
1. Set upProject or Gem instructionsThe rules of the session
2. Song cardStory, details, form, banned wordsThe idea and the point of view
3. Title and chorusTitle options, then options for one lineThe final wording
4. VersesOutlines, then a line check of your draftThe order of events
5. Chords and formProgressions as text, a structureWhat you choose after listening
6. MelodyNothing: you hum and recordThe melody and the voice

What ChatGPT, Gemini and Claude can do for a song

All three are chat assistants that answer in text. For a song, that means they are good at fast options: titles, rhyme lists, alternative lines, outlines, chord progressions written as names or numbers, and section plans. They are weak at three things you need: they can’t hear your melody, so a line that looks right on the page may not fit the notes; their syllable counts are guesses; and with no details they fall back on the most familiar phrasing, which is why “AI lyrics” sound alike.

Here is where each one keeps the context of your song, checked on official pages on October 2, 2026:

AssistantFree planWhere the song card livesWhat you get backPaid from
ChatGPTYes, with usage limitsA project: instructions and files together (5 files per project on Free)TextGo $8/month, Plus $20/month (US)
GeminiYes, with usage limits; Gems includedA Gem: a name, instructions and “Knowledge” filesText, plus music tracks made with Lyria (18+, with limits)Google AI Plus, Pro and Ultra, priced by country
ClaudeYes, with usage limits; up to 5 projectsA projectTextPro $17/month billed yearly or $20 month to month

If you are still choosing between them, our comparison of AI tools for lyrics goes through what each does best and what their terms say.

Step 1: set up the chat before the first prompt

One song, one place. In ChatGPT, create a project named after the working title; in Gemini, make a Gem (Gems, then New Gem, a name and instructions); in Claude, a project. Paste these rules as the instructions, so you don’t have to repeat them in every message:

Prompt 1 (project or Gem instructions): You are my co-writing assistant for one song. Don’t write a full lyric unless I ask for it. Give options as numbered lists. When I paste a draft, keep my wording and only touch what I point to. Flag clichés and abstract words. Never imitate the lyrics of a real artist. Write in US English. Keep answers short.

Before you paste unreleased lyrics, check the privacy settings. In ChatGPT, turning off “Improve the model for everyone” in Settings, Data controls keeps new conversations out of model training, and a Temporary Chat is not used for training either. In Gemini, chats with Keep Activity off, and temporary chats, are not used to train Google’s models; Google also asks you not to enter anything confidential, because a subset of chats is read by human reviewers. In Claude, the model-training toggle is in Settings, Privacy.

Step 2: write the song card (copy this template)

The song card is the brief you would give a co-writer. It is the difference between an average answer and a useful one. Fill it in and send it as your first message, or attach it to the project as a file.

Prompt 2 (song card):
Song: [one sentence about what happens]
Who sings, to whom: [I, to …]
Where and when: [a place and a moment]
Real details: [two or three objects, sounds or words someone said]
Style: [genre, tempo, one reference for tone, not for copying]
Form: [lines per section, syllables per line, which lines rhyme]
Banned words: [heart, soul, fire, rain, tears, forever, plus yours]
Reply only “Got it” and wait for my next request.

Here is the card for the worked example on this page. It’s an original song, and every answer quoted below is an illustration I wrote to show the choices, not the output of a particular tool:

Song: Every Sunday I call my sister from the laundromat, the only hour of my week with no rush.
Who sings, to whom: I, to my older sister back home.
Where and when: a laundromat in the city, Sunday evening, two years after I moved.
Real details: the dryer only takes quarters; her dog barks at the mailman; she always asks if I’m eating.
Style: folk-pop, about 92 BPM, acoustic guitar.
Form: 4-line verses and chorus, about 7–8 syllables a line, lines 2 and 4 rhyme, title in line 1.
Banned words: heart, soul, fire, rain, tears, forever, journey, miles apart.

Step 3: get the title and the chorus first

The chorus carries the title and gets sung most, so it comes first. Ask for titles, not lyrics:

Prompt 3 (titles): From the song card, give me 10 titles of two to five words. Use only the details I gave you, and include phrases someone in this story would say out loud.

From a list like “Quarters for the Dryer”, “Are You Eating?”, “Sunday Spin”, “Forty Minutes” and “Low Heat”, I picked “Forty Minutes”: it’s the length of a dryer cycle and of the call, so the title holds the whole idea. If you would rather start from a random spark, the free song title generator works without a chat.

Now write a rough chorus yourself, even a clumsy one, and ask about one line only:

Forty minutes on the dryer,
forty minutes on the phone,
you keep asking if I’m eating,
and I miss you, I’m alone.

Line 4 names the feeling and rhymes on an obvious word. The narrow request:

Prompt 4 (one line): Here is my chorus: [chorus]. Line 4 names the feeling. Give me 8 options for line 4 only, 7 syllables, ending on a perfect or near rhyme with “phone”. Show it through something the narrator does or hears. Don’t touch lines 1–3.

An answer of that kind might include “and I tell you I’m not alone”, “like a kid out on his own”, “and I say I’ve finally grown” and “till the buzzer brings me home”. The first three tell you how to feel. The last one has an idea you didn’t have: the dryer’s buzzer ends the call. Keep the idea, write the line in your own words, and check it:

Forty minutes on the dryer,
forty minutes on the phone,
you keep asking if I’m eating
till the buzzer sends me home.

“Home” against “phone” is a near rhyme (same vowel, different final consonant), and it sounds more natural than forcing “alone”. If you want the full list of options before you ask the assistant, the free rhyming dictionary separates perfect and near rhymes. A rhyme request that keeps the assistant from writing lines for you:

Prompt 5 (rhymes): List perfect rhymes and near rhymes for “[word]” in two columns, one or two syllables, only words this narrator would use. Don’t write lines.

Step 4: write the verses one at a time

Never ask for “verse 1 and verse 2” in one go: you get two verses that repeat each other. Ask for the events, then write.

Prompt 6 (verse outline): Don’t write lyrics yet. Give me 3 one-sentence outlines for verse 1: what we see, what happens, what changes. Use only details from the song card.

I took “the call starts: coins on the counter, her dog barking far away” and wrote:

Quarters stacked along the counter,
someone’s sneakers thumping round,
your dog’s barking at the mailman
in our quiet end of town.

Then ask for a reader, not a rewrite:

Prompt 7 (line check): Here is my verse: [verse]. Don’t rewrite it. List any line that is a cliché, abstract or bent to fit a rhyme, with one reason each. Then tell me which lines break the 8-7-8-7 syllable pattern.

Treat the syllable part as a guess, because the assistant reads the words and doesn’t hear them. Paste both the chorus and the verse into the free lyric syllable counter and keep matching lines within one syllable. If the check calls “quiet end of town” common, decide for yourself: I kept it, because the quiet back home is the opposite of the noisy laundromat.

Verse 2 has to move time forward:

Prompt 8 (verse 2): Verse 2 happens [six months later]. Give me 3 outlines where [the call] has changed, so the chorus means something slightly different the second time. No lyrics.

Step 5: chords and structure, as text you then hear

The assistant can suggest chords as names or Roman numerals, but you decide by ear:

Prompt 9 (chords and form): Suggest 3 chord progressions in C major for the verse and 3 for the chorus, as Roman numerals and chord names, around 92 BPM. The verse should feel more held back than the chorus. Then give me a structure under 3:30 with the length of each section in bars.

Don’t trust the progression until you hear it. For this song I tried the verse on vi–IV–I–V (Am–F–C–G in the chord player) and the chorus on I–V–vi–IV (C–G–Am–F with drums). For the form, lay out the sections in the free song structure planner: a classic verse, chorus, verse, chorus, bridge, chorus fits easily in three and a half minutes at 92 BPM.

Step 6: make the melody yourself (hum, record, choose)

This is where your song stops being a chat. ChatGPT can’t hear you, and that is fine: the melody is the part people remember, and it should be yours.

  1. Loop the chorus chords in the chord player and hum over them three times without words. Change something each time: start higher, start later, hold the first note longer.
  2. Record every pass. The free online voice recorder works in the browser and the audio never leaves your device.
  3. Sing the words onto the best hum. Put the stressed syllables (FOR-ty MIN-utes) on the strong beats, and give the title the longest notes.
  4. If the chorus feels too high or too low, find a key that fits your range with what key should I sing in, then move the chord loop to that key.
  5. Keep a steady tempo while you practice with the tap tempo metronome.

If you can’t play an instrument, this is still a complete method; the guide to writing a song without an instrument goes further with a 7-day plan.

When Suno or Gemini make the audio: what you trade

Many people searching “ChatGPT song maker” want to hear their lyric sung. A text answer can’t do that; a music generator can, and Gemini now has one built in. The two you are most likely to meet, checked on October 2, 2026:

  • Suno turns a prompt or your own lyric into a full song with vocals. The free plan gives 50 credits a day, with no song downloads and no commercial rights; Pro starts at $8 a month billed yearly and includes commercial use. Its terms (effective September 3, 2026) assign rights in the output to paying subscribers but state that Suno makes no promise any copyright will exist in it.
  • Gemini can now make a music track in the same app with Google’s Lyria model: a short one of about a minute or a full one of two to three minutes. Google says limits apply, you must be 18 or over with Keep Activity on, and every track carries an invisible SynthID watermark. Longer tracks reached paid subscribers first.

The trade-off is the same with both: the melody, the arrangement and the voice you hear come from the model, not from you. That is useful as a quick reference to hear how a lyric might sit, and a problem if you want to sing the song as yours or release it. If you want other ways to set your words to music, with or without AI, see these apps to put music to lyrics.

Is a song written with ChatGPT yours? Originality and rights

OpenAI’s terms say that you own the output, as far as the law allows, and they also warn that output may not be unique: other users can get similar text. So check originality before a lyric leaves your notebook. Search any line that sounds too smooth in quotation marks, never paste someone else’s lyrics to imitate, and don’t ask for a song “in the style of” an artist’s lyrics.

The law is a separate question from the terms. In the US, the Copyright Office concluded in January 2025 that AI output is protected only where a human author has determined enough of the expressive elements, and that prompts alone are not enough. This is not legal advice, but the practical rule follows: the more of the final words, the melody and the choices are yours, the stronger your claim. Keep your drafts and takes as a record of what you wrote. Our guide to copyright for AI music covers the US, UK and Canada in detail.

Mistakes that make a ChatGPT song sound generic

  • Asking for the whole song. You get the average of every song on the subject. Ask for one small job per message.
  • Taking the first answer. Ask for eight options and pick the one with an idea you didn’t have, then rewrite it.
  • Correcting in long threads. After five corrections the chat drifts. Edit your earlier message or open a new chat in the same project with the song card.
  • Letting it pick the rhyme scheme. Set the scheme in the song card, or every line will end on fire and desire.
  • Trusting its syllable counts. Count with a tool and sing the line before you believe it.
  • Skipping the melody. A lyric is half a song. Hum it before you polish it.

The full prompt sheet to copy

Paste this into a note and replace the brackets for each new song:

1 SETUP: You are my co-writing assistant for one song. Don't write a full lyric unless I ask. Numbered options only. Keep my wording; touch only what I point to. Flag clichés. Never imitate a real artist's lyrics.
2 SONG CARD: Song / Who sings, to whom / Where and when / Real details / Style / Form / Banned words. Reply "Got it".
3 TITLES: 10 titles of 2–5 words from the card's details, things someone in this story would say.
4 ONE LINE: Here is my [section]. Give 8 options for line [n] only, [x] syllables, rhyming with "[word]". Show, don't name the feeling.
5 RHYMES: Perfect and near rhymes for "[word]" in two columns. No lines.
6 OUTLINE: 3 one-sentence outlines for [verse 1]: what we see, what happens, what changes. No lyrics.
7 LINE CHECK: Don't rewrite. List clichés, abstract lines and rhyme-bent lines, one reason each. Flag lines that break [8-7-8-7].
8 VERSE 2: Verse 2 happens [later]. 3 outlines where [detail] has changed and the chorus means something new.
9 CHORDS AND FORM: 3 progressions for verse and chorus in [key], Roman numerals and names, [BPM]. A structure under [3:30] in bars.
10 FINAL READ: Read the full lyric as an editor. List repeated words, tense changes and point-of-view slips. Don't rewrite.

Where Zoundroom fits: the AI options next to the song itself

A ChatGPT session leaves the lyric in a chat history, the hums in your phone’s recorder and the chords in a browser tab. In Zoundroom, each song is a project with its recordings, lyrics, notes and files, on iOS, Android and the web, and Auralis, the assistant inside the app, is the in-app alternative to the prompts above: it writes text only (lyric ideas, alternative lines, chords written out and structures), never audio, and it doesn’t listen to your recordings. It takes into account what you tell it about your music in your Creative DNA. The moment it helps most is the one in steps 3 and 4: write the chorus in a note, ask Auralis for options on the weak line, rewrite the one you keep, then record yourself singing it inside the same note, where the audio lands at the cursor, so the draft, the choice and the take stay together. Be clear about the limits: Zoundroom doesn’t generate audio, has no virtual instruments or beats, and isn’t where you mix and master a finished track (a DAW such as GarageBand or BandLab is). Next step: create a project with your working title, paste your song card into its first note and record your first hum of the chorus under it.

Frequently asked questions

Can ChatGPT make a song with music and vocals?

In a chat about lyrics, ChatGPT gives you text: the lyric, chords written out and a structure. To hear a sung track you need a music generator such as Suno, or the Gemini app’s music feature. Or sing it yourself over a chord loop, which keeps the melody yours.

What is a good ChatGPT prompt for song lyrics?

One that gives details and a narrow job. Send a song card first (story, who sings to whom, real details, form, banned words), then ask for options for one line or ten titles instead of a full song. The ten prompts above cover a whole session.

Is it free to write a song with ChatGPT or Gemini?

Yes, both have free plans with usage limits, and so does Claude. Paid plans add more usage: ChatGPT Go costs $8 a month and Plus $20 in the US. Suno’s free plan doesn’t allow downloads or commercial use.

Can I use a song written with ChatGPT commercially?

OpenAI’s terms give you ownership of the output as far as the law allows, but copyright protection depends on how much of the song you wrote. Rewrite the lines in your own words, make the melody yourself and keep your drafts. For a release, read the terms of every tool you used.

Which is better for songwriting, ChatGPT or Gemini?

For lyric text they do the same job; the difference is where you work. ChatGPT projects and Gemini Gems both keep the song card for you. Gemini can also generate a music track, which is useful as a reference but isn’t your melody.

Sources

Official pages checked on October 2, 2026: