reliability question: how do you handle llama.cpp?

I am trying to improve my approach to llama.cpp and could use some practical advice.

I am particularly interested in reliability and advanced workflows. I would like to compare approaches that people have actually used rather than just theoretical suggestions.

What would you tell someone trying this for the first time?