I'm not against the prompt being changed, the point I was making is that an LLM is prone to the exact same mistakes even if you change the prompt. A trivial example is the very basic character counting mistake, I just asked chatgpt:

> How many p are in strawperry?

> There are 0 “p”s in strawperry.

And I can trigger the same mistake with various words even when adjusting the prompt many times. So I cannot teach chatgpt to correctly count characters.