Triage for weak results: naming the problem before fixing it

'It's not good' is not a diagnosis

The moment a generation disappoints, most people reach for regenerate. That's a coin flip dressed as a decision. Before you spend another run, spend ninety seconds identifying which layer is actually failing, because the fix for a structural problem and the fix for a tonal one have nothing in common.

Listen once more with a single question in mind: at what moment did I lose interest? The answer usually points straight at the layer.

Be precise about the moment. 'The second verse' is a usable answer; 'the middle bit' is not. If you can name a timestamp, you can usually name the layer within a few seconds of thinking about what changed there — or, more often, what failed to change there.

Locate the layer

The contradiction sweep

Before anything else, reread your prompt looking only for words that fight. This takes a minute and resolves a large share of erratic results.

• Two genres with no stated host.
• Tempo stated that doesn't match the genre's normal range.
• Vocal descriptors from incompatible slots (whispered and belted).
• Space and intimacy asked for simultaneously.
• Tone words pulling opposite directions (dark and crisp).
• More than six named instruments with no roles.
• Adjectives that describe a feeling rather than a sound.

Fix in order, never in parallel

• Structure first — nothing else matters if the song doesn't go anywhere.
• Then arrangement — decide who plays what and when.
• Then vocal — the voice sits on top of a working arrangement.
• Then production — space, width and tone, last.
• One change per run, two runs per change, throughout.

Fixing bottom-up feels slower and is dramatically faster, because each layer constrains the next. Polishing the production of a track with no dynamic arc is painting a wall that's about to be knocked down.

The ordering also protects your judgement. When you change one layer at a time you always know what caused an improvement; when you change three, you'll attribute the win to whichever change you're most attached to, and carry a superstition into every future prompt.

When the idea is the problem

The hardest diagnosis is the one where nothing is technically wrong. The mix is clean, the vocal is fine, the structure is textbook, and the song is inert. That's not a prompt bug — it's an absence of a point of view.

Prompts that produce interesting music usually contain a specific tension: a contrast between two elements, a mood that doesn't match its genre, a constraint that forces a choice. 'Upbeat pop song' has no tension. 'Cheerful melody, devastating lyric, arrangement thins out at the end' does. Add one deliberate contradiction — an intentional one, unlike the accidental kind in the sweep above — and the same technical settings produce something worth hearing.

Flat:    dark ambient track, atmospheric, moody
Better:  dark ambient built from a single detuned music-box loop;
         warm tape hiss underneath; one moment at 2:00 where everything
         drops to silence for two bars

Knowing when to stop

• Three changes with no audible improvement — the concept has a ceiling, start over.
• You're fixing things faster done in an editor — export and edit.
• You've lost track of what you changed — go back to the last logged version.
• The track is fine and you're polishing from anxiety — export it.
• You've stopped enjoying it — leave it a day; your ears are the instrument and they get tired.

FAQ