I experimented with prompting Suno and noticed something interesting: the model responds far more strongly to production primitives (808 slides, triplet hi-hats, etc.) than descriptive prompts.

The post walks through some prompts I tested and includes generated tracks so you can hear the differences.