Tests
A Repeatable Prompt Testing Plan on Create Music
Run Create Music like an experiment: one hypothesis, a saved baseline, a short scorecard, and a clear keep/revert rule.
Cette astuce est rédigée en anglais. D’autres langues arriveront.
Ask one question before you generate
Good tests answer a single audible question: “Does a drier vocal sit clearer on this chorus?” Bad tests change Style, Lyrics, and tempo language together and then guess.
Write the question in your notes next to the version number. If you cannot state it, you are exploring — budget credits accordingly.
Save a baseline you can return to
Copy the exact Style string and Lyrics outline from the control take. Restoring the baseline is how you know a change helped. Without it, every “improvement” is a new random walk.
Change one decision category, not one random word
Categories mirror the Create Music lanes: genre home, groove, instrument roles, vocal knobs, production space, or Lyrics section contrast. Change one category per run so the listening result maps to a cause.
Use a four-point listening scorecard
Score quickly (1–5): hook clarity, vocal fit, section contrast, muddiness (invert so higher is better). Listen in three passes — chorus only, verse only, full mix at one volume.
Keep / revert / narrow: keep if two scores rise and none fall; revert if muddiness worsens; narrow if only one section improved.
Build a tiny evidence library
Store winning Style fragments (“dry forward vocal”, “chorus percussion lift”) with the song context that proved them. Reuse fragments across projects instead of reinventing adjectives.
Stop rule: three failed tests on the same question → rewrite the listening goal or switch tools (Extend a good section, Persona for voice drift, Reference for production target).