MUSIC

Most Suno advice is about prompting. Words, genres, style strings, the magic bracketed incantations that supposedly unlock a better song. Having made a pile of tracks I can tell you the prompt matters less than people think, and the parts that actually decide whether a track survives are almost never discussed.

Generation is cheap. Selection is the work.

A prompt gives you four candidates in about a minute. That is the easy 5%. The other 95% is deciding which one is real, and almost every beginner keeps the first clip that sounds impressive on the first listen.

First-listen impressiveness is mostly novelty. A track can sound enormous on first play and be unlistenable on the third, because the thing that grabbed you was one section and the rest is filler that repeats it. Here is the test I use:

• Listen to the intro twice in a row. If it feels slower the second time, the arrangement is thin.
• Find the moment you would skip. There is usually one. That is the section that needs to not exist.
• Check the last 20 seconds. Generated tracks frequently run out of ideas and resolve by repeating the hook slightly quieter. If the ending is a fade of the middle, the track has no ending.

Structure beats style

The biggest quality jump in my own output came from specifying structure explicitly rather than describing a genre. Two prompts with the same style string, one of which names where the change happens, produce wildly different results.

What I have found works:

[Intro]     sparse, single instrument, no drums
[Verse]     drums enter, restrained vocal
[Pre]       build, remove the bass
[Chorus]    full arrangement
[Break]     drop everything for four bars
[Chorus]    return, add one new layer
[Outro]     strip to the intro instrument

That template alone fixes the two most common failures: a chorus that arrives with nothing to contrast against, and an ending that does not exist. The break is doing the heaviest lifting: pulling everything out for four bars makes the return feel twice as big for almost no extra content.

Iteration recipes that actually change the result

Persona consistency across tracks

If you want a set of songs that sound like one artist, pin a style string you reuse exactly and vary only the structure and lyrics. Changing the style string between tracks is the fastest way to sound like a compilation of two different people.

Fixing a good-but-wrong track

When a track is 80% right and one section is wrong, do not regenerate the whole thing hoping for better luck. Use the extend or cover features to replace just that section, then splice. It feels like more work and it is dramatically faster.

Lyrics as arrangement

Line length shapes rhythm. A chorus of eight short lines will always sound rushed at a given tempo; four longer lines will breathe. Before blaming the generation, check whether the lyric is fighting the bar.

On keeping the interesting failures

Some of the best things I have made came from a generation that misread the prompt. A lo-fi prompt that came back as glitch-hop. A simple groove that arrived oddly detuned.

Those are not mistakes to be corrected. They are the only part of the process where something arrives that you would not have asked for. And that is the closest an AI-assisted workflow gets to surprise. When one shows up, keep it, then build the song around what it accidentally did.

The workflow end to end

1. Write the structure first, as a skeleton with the sections named.
2. Pick a style string and reuse it verbatim across a session.
3. Generate four takes. Listen to the last 20 seconds of each first, the endings sort good from bad faster than anything else.
4. Keep at most one. Replace broken sections rather than re-rolling.
5. Leave it alone for a day. Listen once more. Most tracks do not survive that.

The tracks that do survive are on the music page. That page reads straight from the published catalogue, so it is always whatever is currently live rather than a hand-maintained list.