Spent an hour writing a long list of tone rules once when three example emails would have done the job faster. Adding one bad example on purpose is smart though, since I usually just sit there hoping it reads my mind.
The "3 fake customer emails as examples" part is the key here and people sleep on it. You basically handed the bot a tiny map of what good looks like instead of writing a page of rules about tone and phrasing. I did the same thing when mine kept replying like a robot, dropped in a few sample replies I actually liked, and it picked up the pattern way faster than any instruction I typed out. What most people miss is you should mix in a bad example too, because just showing good ones leaves it guessing where the line is. Throw in one reply that is too stiff or too pushy and tell it not to do that, and the whole thing snaps into place. Three is honestly the magic number, any more and it starts blending them into mush instead of learning the vibe.