An API is good when you do not think about it

You should always Implement what is conceptually correct first. Once that is done you should implement the user-interface with predictability in mind. To do this you always need to ask yourself: "is that behavior surprising?" "what intent are these names conveying?".

An API is an artifical language on top of an artificial language (the one with which it is implemented), and should come as close as possible to a human language. So interacting with your code activates the more or less the same neural pathways as the ones you use when thinking about the problem.

Sometimes the API will match exactly the problem space, but sometimes predictability will come from re-using idioms that people are used to. For instance JAX took off quickly because it made the right call by copying the numpy user interface: one just had to replace numpy as np with jax.numpy as jnp to get the same operations on DeviceArray that one has on ndarray.