Making Original Music With AI: Workflow, Limits and Taste

AI music generators are extremely good at producing something listenable and extremely bad at producing something memorable. The difference is almost entirely in what happens before and after generation. Here is how a Faithsmove track actually gets made.

Write the brief before you touch a generator

Every track starts as four lines in a notebook: what it is for, who it speaks to, the feeling in one word, and the reference feel — usually described in instruments and tempo rather than by naming an artist to imitate. "Ninety-two BPM, warm Rhodes, brushed drums, one voice, hopeful but tired" is a brief. "Afrobeats banger" is a wish.

If the track scores a film, the brief inherits the beat sheet: where the emotional turn happens, at what timecode, and what the music should be doing on either side of it.

Structure first, sound second

I decide the arrangement on paper — intro, verse, lift, chorus, break, final chorus, outro — with bar counts. Generators drift structurally: they will loop pleasantly forever and never build. Knowing where the lift lands means I can generate sections separately and assemble them, rather than hoping a single generation delivers a shape.

Generate in parts, keep the alternates

Rather than one full track, I generate stems and sections: a chord bed, a rhythm idea, a hook, a bridge. Several takes each. Then I audition them against the brief, not against each other — the prettiest loop often serves the song worst. Alternates get kept; a discarded bridge from track two frequently becomes the intro of track five.

Vocals are where taste shows

Generated vocals betray themselves in three places: breath, timing and consonants. So I write lyrics that give the voice somewhere to breathe, avoid dense syllable runs, and prefer open vowels on sustained notes. Where a line refuses to sit naturally, I rewrite the line rather than push the model harder. Doubling, gentle pitch correction, and a real sense of pocket in the timing do more for believability than any "realistic vocal" prompt.

Mix it like real audio, because it is

The most common reason AI tracks sound cheap is that nobody mixed them. Generated material arrives loud, mid-heavy and glued together in the wrong way. My default pass:

• High-pass everything that is not bass or kick, so the low end has room.
• Carve a small dip around 300–500 Hz where generated mixes tend to pile up.
• Give the lead vocal its own space with subtle sidechain ducking on the pads.
• Use one reverb space for the whole track so separately generated parts sound like one room.
• Automate arrangement dynamics by hand — drop elements before the chorus, because generators rarely will.

Master gently. Loudness will not rescue a song that has no dynamic shape, and squashed AI material collapses fast.

Check it where people will hear it

Phone speaker, cheap earbuds, car. If the hook does not survive a phone speaker, the arrangement is too dense. This check has changed more of my mixes than any plugin.

The lines I do not cross

I do not clone identifiable artists' voices, and I do not prompt "in the style of" a named living performer to pass work off as theirs. Influence is legitimate; impersonation is not. Every release is disclosed as AI-assisted, lyrics are written by hand, and where a human performer contributes, they are credited and paid. These constraints cost very little creatively and protect everything else.

What AI actually changed

It did not remove the need for taste — it removed the wait. Ideas that used to require a session musician and a booked studio can be auditioned in an afternoon, which means more ideas get tested and worse ones get discarded early. The song still has to be written, arranged, mixed and judged by a person. That part has not been automated, and honestly, that is the part worth doing.