I only have very limited experience, but from what I've personally figured out, another good practice is to look into the thinking process (even though models can and do "lie" in there, so it's not really the actual thing that's happening under the hood) and catch mistakes early before the AI has wasted a bunch of time and (possibly) money on some nonsensical detour. You then have to stop the process early, add a line or two to the prompt to explicitly stop it from choosing this detour and start again. (In my case - I almost exclusively use local LLMs - it's not about money, at least not directly, but because there's only a very limited context memory due to the amount of RAM I have, so I don't want it to be filled up by stuff that's not helpful.)

It's like babysitting a toddler. The moment you look away, chances are it might try something you really don't want to happen.

Sometimes, the LLM might misunderstand your prompt entirely, because you chose imprecise language, made a typo that changes the meaning, you asked something of it that is way outside of its ballpark/training data, because of missing context or because the AI just had another dumb moment.

The old adage that computers are just as smart or dumb as the user (but much faster at it) often applies to AI as well, at least in certain use cases - in others, AI can actually help less intelligent users, but at the same time drag down smart people and erode their skills over time. It's weird what it's doing to our brains, which is why I'm very wary of using it for anything other than experimentation and fun, and avoid using it for things that are actually important to me. Many people don't have this choice though, with mandatory AI usage in many orgs still being on the rise.