Bloomkeeper will not nail your voice on the first go. What matters is that it gets closer every time you correct it, and correcting it is easy.
When a draft is close but not quite right, edit it in the approval screen and then approve it. Bloomkeeper keeps the difference between what it wrote and what you changed it to, and treats that as an instruction.
Cut the word "delighted" three times and it stops writing "delighted". Shorten its long sentences a few times and its sentences get shorter. The corrections build up.
This is much more effective than rejecting a draft and asking for another, because a rejection only says no, while an edit says what you wanted instead.
If something is badly wrong rather than slightly off, reject it and say briefly what was wrong. Bloomkeeper writes it again with that in mind.
Bloomkeeper will not write "curated", "nestled", "look no further", "reach out" or "we pride ourselves". It does not open with "In today`s world", and it does not tell your customers what they supposedly already know. Your own examples and corrections always win over these defaults.
If drafts are consistently wrong rather than occasionally wrong, the problem is usually the material it learned from. Go back to Brand voice and paste in more real customer messages. Five good ones beat fifty product pages written by someone else.