This “dopamine hit” isn’t a permanent source of happiness, just repeatedly clicking “randomize” button not going to make you feel constantly high, after 3 maybe 5 hits you will start noticing a common pattern that gets old really fast. And to make it better you need to come up with ways to declare different structures, to establish rulesets, checklists, to make some unique pieces at certain checkpoints yourself, while allowing LLM to fill all the boilerplate around it, etc. Which is more effort but also produces more rewarding results. I like to think about it this way: LLM produces the best most generic thing possible for the prompt. Then I look at it and consider which parts I want to be less generic and reprompt. In programming or scripting, I’m okay with “best generic thing” that solves the problem I have. If I were writing novels, maybe it’s usable for some kind of top-down writing where you start with high-level structure, then clarify it step by step down to the lowest level. You can use AI to write around this structure, and if something is too boring/generic it’s again simply a matter of refining this structure more and expanding something into multiple more detailed things.