Different LLMs respond to requests slightly differently. So if I say - "here's a log of all the errors I got, can you help me solve it?" - you will get different suggestions than if provided more constraints first. It helps to split the logs and ask to solve one problem at a time, rather than expect it to solve all of the problems at once. So it's easy to just ask for help, but you get better results by helping the LLM with extract directives first. That's one example.

If you ask o1 vs Grok to write something, they have very different styles. Grok is chill and concise, o1 is professory and kind of dull. But, each has its advantage, so it makes sense to give o1 something to make it better, then feed to Grok to make it more fun. Things like that.