← Back
EmergingIndie 5 min read

AI as Co-Writer: The Two-Hour Sprint

A structured two-hour songwriting session using AI assistance — for the unblocking, not the writing.

by · · 5 min read release-strategy

AI co-writing works when you treat the model the way a good session player works — it brings options to the room, you make the call. A two-hour sprint is the sweet spot. Long enough to finish a verse-chorus draft, short enough that the energy and judgment stay in your hands instead of slowly drifting to the screen.

The shape of the two hours

The structure matters more than the tool. Most failed AI co-writing sessions fail because the writer opens the chat with no plan and lets the model shape the agenda. Reverse it.

  1. 0:00–0:20

    Premise, by hand

    Voice memo or notebook. What is this song about, in one sentence? Who is talking? Who are they talking to? No AI yet — the meaning of the song has to come from you.

  2. 0:20–1:00

    Draft + alternates

    Write the verse and chorus rough. Use ChatGPT or Claude only for rhyme alternates, image variants, and the question 'is this melody phrase too crowded for the syllable count?'

  3. 1:00–1:30

    Sanity-check the structure

    Paste the draft into the model and ask: where does the energy drop, what's the weakest line, what's the strongest. Take notes; do not accept rewrites verbatim.

  4. 1:30–2:00

    Voice memo + final pass

    Record yourself singing it acapella. Anything you stumble over, fix by hand. Close the laptop before you hit save.

Two hours, four blocks. Phone face-down. Tool open only when it earns its way in.

Use AI for the parts that don't matter so you can spend your energy on the parts that do. The meaning of the song is the part that does.

What AI is good at, what it isn't

AI is good at the friction work. Rhyme alternates, syllable counts, "give me fifteen ways to phrase this without using the word night." Structural sanity checks. Counter-arguments for lyrics that are too on-the-nose. Suno and Udio are useful as reference-track generators — a quick way to hear roughly how a tempo or feel would sit — but treat their output as a sketch, never as a part of the record.

AI is bad at the parts of songwriting that matter most. It cannot tell you what the song is actually about. It cannot decide whether the hook should sit on the four or the four-and. It cannot weigh the difference between a line that's clever and a line that's true.

The never-AI rule

Pick one element of every song that never touches the model. The opening image. The bridge concept. The hook itself. That element is where your voice lives, and protecting it from the model is what keeps the catalog from sounding interchangeable five years from now.

The sprint ends with a voice memo. If you can sing it acapella and the song still feels like yours, it is. If it sounds like anyone could've written it, the model wrote too much of it. Throw out the lines that aren't yours and try again tomorrow.

Frequently asked

Should I tell anyone the song used AI in the process?
Only if AI wrote a top-line melody or finished lyrics that ended up on the record — in which case yes, both ethically and for split-sheet reasons. Using it for brainstorming, rhyme alternates, or structural notes is the same category as using a thesaurus and doesn't require disclosure. The line is whether AI authored a part of the final work.
Which tool — ChatGPT, Claude, Suno, Udio?
ChatGPT and Claude for lyric and structure work — they're the better co-writer surface because they push back and offer alternatives. Suno and Udio are for reference tracks and feel-checks, not for finished material. Most working songwriters we talk to settle on one chat model for words and never let a generative-music tool render the actual song.
What if AI gives me a line that's actually great?
Rewrite it in your own voice before it goes in. The test is whether you could sing it acapella to a stranger and explain why that exact phrasing matters. If you can't, the line isn't yours yet — keep working it. The line in the final song should sound like you, not like a model.
Can a two-hour sprint really finish a song?
A verse-chorus draft, yes. A finished, mixed song, no. The sprint targets the unfinished-idea-to-rough-demo gap — the place where most songs die in a notes app. Polish, edits, and the bridge are separate sessions, ideally with at least a day between them.