I think the trick with LLMs is realising that all they do is repeatedly randomly choose a word to output from a weighted set, and the weights are influenced by the context of the preceding words. Your role is to set up that context so that it outputs the words that are most useful to you.

That's why ChatGPT has the Custom Instructions feature. Put in there (or up front in the conversation) that you want it to consider whether what you are asking for is possible or not, and if there are alternatives it could provide. Then work with it collaboratively to adjust its context while chatting on a topic (e.g ask it to check its work and other open ended questions) and you'll likely get much more useful results