One principle I try to apply (when possible) comes from when I learned Haskell. Try to keep the low-level logical computations of your program pure, stateless functions. If their inputs are the same, they should always yield the same result. Then pass the results up to the higher level and perform your stateful transformations there.

An example would be: do I/O at the high level (file, network, database I/O), and only do very simple data transformations at these levels (avoid it altogether if possible). Then do the majority of the computational logic in lower level, modular components that have no external side effects. Also, pass all the data around using read-only records (example: Python dataclasses with frozen=True) so you know that nothing is being mutated between these modules.

This boundary generally makes it easier to test computational logic separately from stateful logic. It doesn't work all the time, but it's very helpful in making it easier to understand programs when you can structure programs this way.