Build LLM REPLs into your program itself

Build in interactive LLM loops into any program you run manually that deals with LLM output.

Once this clicks for you, it's like the same feeling you get when you understand recursion for the first time.

I have saved countless hours by --- instead of blindly re-running my programs until the LLM output it generates (whether text or images) is satisfactory --- by building an LLM REPL into the program itself.

For example, I have program that generates quiz questions for me based on the contents of a database. Most of the time, they're great. Sometimes, they're stupid. Instead of repeatedly retrying the stupid ones, my program asks me for additional context (literally lets me type into the terminal and tell the LLM why their generated quiz question was dumb) and from there it generates good content very quickly.