"[" EXPR "|" STMT["," STMT]* "]"
Where STMT is given by: STMT ::= EXPR | PAT "<-" EXPR
The problem is that, until you see that "<-", you don't know whether you're supposed to be parsing a pattern PAT or an expression EXPR. This is extra hard because patterns and expressions overlap significantly; for example, "(x, Left 2)" could be either a pattern or expression. But "(x, Left (2+3))" is definitely an expression. You can get arbitrarily deep into parsing a pattern before you realize it's actually an expression!Recursive descent parsers, even with precedence climbing, can't really handle this nicely. Neither can LL or LR parsers. In fact, GHC's actual parser uses a hack: it parses patterns as expressions, and as a post-processing step checks that the expression it parsed actually was a valid pattern! This works IF your patterns are a subset of your expressions. But if there are patterns that aren't valid expressions (Racket's match-patterns have this property), then you need to get even cleverer.
General-purpose parsing algorithms like GLR, GLL, parsing with derivatives, and Earley parsing (which Kegler advocates) can handle this, of course, although I'm not sure how efficiently they do it.