I used to work on a CASE tool early in my career. I was always struck by the fact that we never ate our own dog food; that is, we never employed the tool we were working on to guide its own construction. This was the first time my eyes were opened to the fact that perhaps there was a lot of junk being built :)
CASE itself I believe was an attempt to replicate for assembling bits some of the benefits CAD/CAM was bringing to assembling atoms.
Although the latter is partly because the idea of doing 'SE' without the 'CA' is now inconcievable, even to senior management.
Same about dogfooding. When I realized RSA wasn't (at least when I was working on it) developed with RSA or any UML I felt weird... Doubly weird since other tools like IBM Jazz (team collaboration) used themselves to ensure a certain dose of value.
I found it to be a pain as a student (it clearly was a fad at that point), but every time I go back to shipped code in mainstream paradigms, I understand why people went this way. It's as boring as intractable. We don't want to code at that level, we want abstractions, but these paradigms didn't help so we needed diagrams. And it's an offspring of OOP mostly which helped reducing cpp-like complexity but not for free.
IMO UML boils down to a visual type checker. We can layout things more efficiently than with linear text, and can appreciate the relationships between parts. But in a way, that's what a type checker does. I remember never needing a diagram (for medium single person programs that is) in Eclipse since it could look at the whole system for me and show contradictions or suggest options.
I'm now very regularly tempted to do things with a bit of empty interface logic checking, FP-ish minded. You can still write this as text, and as long as it the types flow logically it's OK.
ps: this was also influenced a bit by posts like these https://www.google.com/search?q=f+sharp+uml