The problem is that LLM's sometimes gets the right answer and then you are like Wow this is the best! And the next minute you are thinking It must be me not giving enough context? Let me try a different model. which then also fails.