I think the reason its so hard is when we ask something we are missing out huge bits of information about exactly what we want. Another human knows from experience and context what the person actually wants. When I say I want a webstore for selling keyboards A web dev already has a pretty good idea what I want because they know what a web store does and how it functions and they know how keyboards work so they will know to add search filters for things like switch types even though none of this information was in the original request.

Explaining exactly what we want and covering all of the edge cases is just as hard as using a regular programming language. Until we can have full back and conversation with computers that have a deep understanding of what people want from a program and can ask the user questions like "But what should happen when x happens?".

Using natural language for stuff like that would be just like writing legalese, except being even more specific about everything.

That actually sounds way worse than just using queries.

How would you write automated tests for it? How would you keep those tests from being really flakey?

I remember the natural language interface in Q&A by Symantec worked pretty well. I was using it in the very early 1990s.

https://en.wikipedia.org/wiki/Q%26A_(Symantec)

The fact that it didn't respond to commands as well as Zork or A Mind Forever Voyaging didn't help the case. Infocom had a pretty solid engine for "natural language" input. Too bad they couldn't adapt it well.