9. Common pitfalls
← Song breakdowns · Guide contents · Next: Future experiments →
Treating descriptors as magic words
Section titled “Treating descriptors as magic words”It is tempting to find one generation I like and credit the most unusual phrase in its prompt. My behavioural tests made that difficult to defend. Emergent Melodic Fragments became piano-led beside Continuous Evolution, then appeared as vocal ad-libs when I changed the motion to Drifting Flow.
A descriptor is not a switch with one guaranteed result. I treat it as a musical direction whose meaning is shaped by the rest of the prompt, the source audio, the model and ordinary generative variation.
Assuming more words mean more control
Section titled “Assuming more words mean more control”A longer prompt can feel more controlled because more decisions are visible on the page. That only helps when the new words have distinct jobs. If several descriptors are doing the same thing, I remove the weaker ones. If two instructions compete, I either place them in different layers or decide which one should win.
Giving every idea equal priority
Section titled “Giving every idea equal priority”Pressure would not have worked as an equal vote between vocal DnB and instrumental metal. The DnB source was the song; the metal source supplied the guitar vocabulary. Once that hierarchy was explicit, Mashup had a much clearer problem to solve.
I use Priority in the same way inside a prompt. It protects the idea that cannot be traded away. If everything is equally important, the field makes no decision.
Changing several variables at once
Section titled “Changing several variables at once”I can change the prompt, Weirdness, Style Influence and Audio Influence together and still get a better song. The problem is that I learn very little from it. When I want to understand a change, I hold the other variables steady. Foggy Hearts separated its Behaviour rewrite from its final Confidence adjustment for exactly that reason.
Ignoring the audio Anchor
Section titled “Ignoring the audio Anchor”The Cover prompt does not deserve credit for every quality in the result. Melody, phrasing, structure and emotional identity may already be present in the source. I try to listen for that inheritance before deciding which prompt line was successful. Comparing Cover with a fresh Generate can make the source’s contribution much easier to hear.
Treating current behaviour as permanent
Section titled “Treating current behaviour as permanent”Everything in this edition was made and observed in Suno v5.5. A later model may interpret the same phrase or influence setting differently. When an observation is worth keeping, I record the model and date with it so it remains possible to retest rather than quietly becoming folklore.
Failing to keep notes
Section titled “Failing to keep notes”The useful record is not only the prompt that worked. I save the settings, source link, model and what I heard in the versions that did not work too. A small chain of problem, change and result teaches me more than a large list of successful tags with no context.
← Song breakdowns · Guide contents · Next: Future experiments →