Generally speaking, I work by building some data that represent my understanding of the model I'm working with, and some functions to operate on that model. My models start out simple and more or less wrong, and as I work, I build improved understanding. Mostly my understanding improves incrementally, and so does my code. I build up the app piece by piece, adding and changing things bit by bit. So most of the time, changing a small piece is all I need, and that's what evaluating one expression does.

Moreover, if I recompile everything, then I'm rebuilding my in-memory state from scratch. Mostly, that's not what I want. Mostly I work like a sculptor: I make a rough approximation and then refine it stepwise. I don't want to rebuild from scratch every time I discover something new. I want to keep most of what I've built and change just what needs to be changed.

That said, sometimes a discovery does call for a larger-scale reconstruction. That's when I reload a whole file or, in more extreme cases, reload an entire ASDF system of files. Once in a while, I change enough (or make enough mistakes) that it makes sense to blow the whole thing away--kill the Lisp, restart it, and load the system from scratch. But that's fairly rare.

Mostly I want to plop down my clay and tweak it a little at a time toward my goal, discovering and refining it a little at a time.