Comment by ssl-3

From my own observations with OpenAI's bots, it seems like there's nuanced levels.

"Don't do that" is one level. It's weak, but it is directive. It often gets ignored.

"DON'T DO THAT" is another. It may have stronger impact, but it's not much better -- the enhanced capitalization probably tokenizes about the same as the previous mixed-case command, and seems to get about the same result. It can feel good to HAMMER THAT OUT when frustrated, but the caps don't really seem to add much value even though our intent may for it to be interpreted as very deliberate shouting.

"Don't do that, fuckface" is another. The addition of an emphatic and profane quip of an insult seems to generally improve compliance, and produce less occurrence of the undesired behavior. No extra caps required.