Most small AI products cannot fine-tune a music model every day. That does not mean feedback is useless. Likes, dislikes, downloads, comments, and play behavior can train the product layer: which prompts are recommended, which tracks rise in the library, and which patterns are avoided.
A liked track can reveal more than taste. It can show that a certain instrument combination works, that a tempo range is useful, or that a mood label is clear. A disliked track can reveal unwanted vocals, harsh dissonance, weak genre adherence, or cheap timbres. The prompt database can be adjusted without changing the model weights.
Feedback should be attached to context. A dislike on a vocal track means something different if the user had instrumental mode on. A like on a cafe track may be stronger if the track was downloaded. Comments can identify problems that numbers miss: humming, broken loops, loud hats, or misleading titles.
The weekly improvement loop can be simple: export recent prompts and outcomes, group by mode and genre, identify liked and disliked patterns, update random presets, adjust negative prompt language, and keep a changelog. This is enough to make the system feel alive.
BGMFREE's prompt engine should become a memory of what users actually use. The model creates sound, but the product learns taste.