Discussion on the ll1 list about the dangers of creating application-specific languages. make, of course, is the classic example of something that wasn't really fully planned out, then gradually extended, and now is a mess. Graydon's talk is relevant, too. But I liked this bit, from Dan Sugalski:
What I find most effective (modulo heckling from the LL audience here) is "Larry and Bjarne were both language design pros, and look what they ended up with. And you think you can do better?"