To me the answer to this is we shouldn't expect it to read our minds (Though it probably more or less can), there are always many interpretations of any given line of human language, it's full of ambiguity. A prompt is more or less a specification of functionality, expressed by us using highly ambiguous language, so the best a machine can do to turning that into unambiguous functionality, is to make a stab at interpretation, demonstrate the functionality resulting, and look for our response to check that the implemented functionality is what we expected by the behavioural specification we expressed.
Add to this that humans are very badly connected systemically, we are missing an awful lot of information, unlike them, a lot of the time we don't really know exactly what we need, so the onus is again on the Ai to try to help us establish this also.
I am telling you this as someone tasked many times historically as a Systems Engineer, to ellicit and elaborate system requirements, it's often necessary to establish firstly what the stakeholder really wants and needs, rather than what they expressed.