[retrovm]

That presupposes that you can start with a big pile of bad APIs and somehow incrementally approach something worth having. That's not consistent with my experience. In my experience the innermost APIs, the ones that get written down first, are effectively cast in stone and dictate the quality of the whole product, forever. The first draft of a system is the one that should have all its APIs worked out with a pencil before any code is written. You get these accidental complexity explosions precisely because someone thought it was easy and obvious to write a function returning std::vector<string> (or whatever) and didn't consider from the caller's point of view that making the function take an output iterator might be less complex.

[jacobolus]

You are right. It is typically more effective to write one implementation as a hacky prototype, play with it a bit, throw it away, and then rewrite a production implementation from scratch, so that the mistakes of a design made by someone inexperienced don’t get baked in forever.

Unfortunately there are cultural/psychological factors which often preclude or discourage this method, even if it would save time and produce better code in the medium term than either exhaustive planning up front uninformed by practice OR just iterating the initial broken version to continually meet new requirements.

[BlueTemplar]

Lol, ninjaed, see above.