One of the easiest traps with ChatGPT is getting a brilliant answer once and assuming the job is solved.
Run the same sort of task again tomorrow and it can suddenly be oddly literal, miss the tone or take the work in a completely different direction. That variability is part of using generative AI.
If something needs to be repeatable, keep the good example. Give the AI the source material again. Be explicit about the few things that must not change. And, where the result matters, add a check rather than assuming yesterday’s success will repeat itself.
A good prompt helps. A good example helps more. The main thing is remembering that a brilliant Monday does not automatically guarantee a brilliant Tuesday.