It doesn't though. It just gives a better answer because prompting it that way tends to constrain the output to be more like those training examples where a bet

It doesn't though. It just gives a better answer because prompting it that way tends to constrain the output to be more like those training examples where a better sounding answer was given. In natural language, if a request is given to use more wordiness, what's more likely to follow?