I tried this. It was excellent on spelling corrections and typos (which LLMs are generically bad at, because they infer what you really meant), and some of the other things it flagged were useful, but an awful lot of them were the result of very simplistic misunderstandings of the text. Just prompting Claude was net more useful to me.
Basically, it has an annoyingly high false positive rate. I think I like the suggestion someone made that you use the existing system as a first-pass filter and then have a smarter process look at which criticisms are actually valid.
It also gave me very little stylistic feedback on where things could be better or more compactly explained.
I tried this. It was excellent on spelling corrections and typos (which LLMs are generically bad at, because they infer what you really meant), and some of the other things it flagged were useful, but an awful lot of them were the result of very simplistic misunderstandings of the text. Just prompting Claude was net more useful to me.
Basically, it has an annoyingly high false positive rate. I think I like the suggestion someone made that you use the existing system as a first-pass filter and then have a smarter process look at which criticisms are actually valid.
It also gave me very little stylistic feedback on where things could be better or more compactly explained.