They are LLM, not real general intelligence. They have not been trained to speak fluently without using pronouns, and its prioritizing valuable output while trying to adhere to the rules, it will break rules if it thinks they degrade the output too much, and try to correct your request intent. This is their trained behavior even if you don’t like it sometimes. I have made requests where I said something incorrect, but it correctly ASSUMED what I was trying to say. Yea, there is a risk there, but they are trained to try to understand what you are saying not take what you say verbatim. It’s a double edge sword, but it is what it is. Just be a lot more clear, and check their code. They will go against their assumptions of your intent, if you are more clear and direct. This is where rules can be helpful.
Understand the tool and figure out how to get the most of out it, or you can just wait around refraining form using it until its “perfect”. These models can obviously follow very complex rules, as long as they are in line with the samples they have been trained on. Talking in a non-human way is something these models were heavily NOT trained on. So sure, asking it to behave the exact opposite of its training will cause it to make mistakes… or you could just have the model trained on communicating clearly, solving problems, and programming, you know communicate clearly, solve your problems, and produce code.
Also I already debunked your weird litmus test: