I think this is often bad advice. It's too easy to find documentation of this sort that misses out describing key aspects of how the software actually promises to behave.
For example, you might have a graphics library with lots of functions for creating pens and coordinate systems and so on, all meticulously documented, and then one "render" function with a brief description that tells you what the return type is, but no description of (say) the layering or composition model, because that doesn't belong to any one function.
Instead, I recommend starting by documenting the key behaviour of the program, describing its model of the world and defining terminology, and only then trying to document each function in terms of that model and terminology.