Don't you think "having a concrete idea of what sort of code change / end behavior you're looking for" affects the prompts and LLM output?