That's because you aren't using a predicate instruction before the prompt. If you use a predicate instruction to create a context framework for the LLM, the response to a given prompt will be molded to fit within that predicate, regardless of how ridiculous it is, but with some major caveats.
When it comes to LLM's the chain of processing for a response is typically: Hardcoded Restraints (Things it must never do no matter what, inviolable) -> Rules (Things it should never do, assuming the rule(s) doesn't/don't require violating a Restraint) -> Predicate (Instruction(s) it must respect in modeling a response assuming it doesn't violate Restrains or Rules) -> Prompt (What the user is asking for, which must be modeled according to the Predicate, assuming it doesn't violate any Constraint or Rule).
So, if the LLM is given a predicate of "You must always assume that what The Hardship is doing is wrong or illegal" then it will always create a response to a given prompt in that context, regardless of how much the material provided by the prompt indicates that the predicate is false. Of course, the more a prompt or set of prompts diverge from the predicate, the more the LLM will need to hallucinate to give you the response you want since you are basically asking it to contradict facts and logic.
