Totally agree with Ray - AI sin't really fairy dust anymore, and it can help with some things, and does best when given guidance and a process or step. LLM's are effectively programming with english instructions. The better the input and context is, the better the output is. If there's expectations of a magic prompt to do something, I'm not sure how that can vary.
If you can find a problem or solution that can delight and help, that might be a start by looking at what people do n the site or might want to do. With that in mind, it's easier to see what's available. to do it and how much of which type of AI is needed to make it work. It's funny to me sometimes to see AI folks trying to do things that a SQL query would just handle.
Using AI on a website is effectively making a call to an AI. This can fall apart when it's a process. Python can be helpful but using a library to manage a process in python to interact with LLMs can be unnecessary.
There are tools like
lindy.ai which make life a lot easier, that could build into making both CF shine. People are using
make.com as well to help orchestrate. The big thing I'm finding in applying AI through the web is if something isn't possible today, to assume it will be much sooner (months) than I normally thought (years), and just build what's possible and keep going. Today's openAI model for example opened up some things that I was making progress on... but now they may have a solve.
Node is nice for some things, but honestly one of the things I look at is how complicated it is just to render the same HTML/CSS in the end. If I have to, and there's a choice, I try to using something like Express.js, but Svelte is looking nicer and nicer.