Modern object-oriented Python is less about creating a class for everything and more about controlling complexity. A useful class has a clear responsibility, a small public interface, and dependencies that can be replaced during testing.
Prefer composition when behavior needs to vary. A service that receives a repository and a notifier is easier to change than a deep inheritance tree with hidden side effects. Python protocols can describe the methods an object must provide without forcing every implementation into one class hierarchy.
Dataclasses are useful for explicit data models, especially when configuration or domain values need readable representation and comparison. Keep validation close to the boundary where data enters the system. Once invalid data is rejected, the rest of the application becomes simpler.
Dependency injection does not require a large framework. Pass collaborators into constructors or functions. In tests, provide a fake repository or clock. This avoids patching global state and makes the behavior visible in the code.
Finally, use encapsulation for invariants, not ceremony. If an account balance cannot be negative, put that rule where balance changes occur. Good OOP makes illegal states harder to represent and makes correct behavior easy to test.