The harmonization I've come to for "shells are awesome" and "shells suck" is that interactive usage and non-interactive usage are too different to be
fully covered by one tool (i.e., you can occasionally cross over for small tasks but only very small ones). In particular, it's the error handling. Interactive error handling by a human in the loop and fully engaged vs. error handling in a program are too fundamentally different.
In both cases, for small uses, you can stretch your shell or REPL or programming language over to the other case, but for long term use it's just not practical. After all, in theory many languages have REPLs that could have long since replaced shell, but only a handful of very dedicated people actually replace their shell with a REPL for some language. I've tried, and I can't do it, even with shell support libraries in the relevant language.
The suckiness of UNIX shells is mitigated if you view them as "the tool optimized for interactive usage". And once you let them be that, they really aren't so bad. It's only if you try to force them to be programming languages, especially past a couple dozen lines or when they try to do something other than just run lots of commands pretty much blindly, that they become really bad.
Also complicated quoting is a pain, but as a percentage of the number of commands I run in my shell, they aren't that large, they just loom large in my memory. Run history and really look at what you're doing day in, day out with your shell. Unless you've got a really crazy usecase, there's a lot of things like "ls" followed by a "cd dir" and such. It's part of why I can't get into REPL replacement; replacing "cd tmp" with even "cd('tmp')" is, percentage-wise, a huge increase in keystrokes. (And don't forget, the space is much cheaper than left paren; the space bar is huge and I basically have an entire digit dedicated to it, whereas left paren is two not-on-home keys at the same time.)