I'm starting to come around to that view, especially since shell is an unusually productive REPL. What are the "right circumstances" for a shell script in your experience?
Wrapper functions. If you want to extend or modify the functionality of already existing tools, but without altering them too much, so it wouldn't be worth to be a new tool on its own.
Admittedly, its not always apparent. I think a general rule of thumb for me is if I can effectively write the script in my mind ahead of time. Anything beyond that and I might regret choosing a shell script for the purpose.
Sometimes a shell script just proves a concept out and is then rewritten later in a production language.