I did not like CLOS before I started working on ChangeSafe. It was "too complicated". It has weird things like "effective slots" and "method combinators" and update-instance-for-redefined-class. Most of CLOS seems to consist of circular definitions, and the documentation reflects this as well. When you are a blind man examining an elephant, no matter what you think you found, there seems to be an awful lot of it with no apparent purpose.
Code doesn't spring into life on the first iteration. It often takes many, many attempts at implementing something before you understand the problem you are trying to solve and figure out the best way to approach the solution. With iterative development, you re-approach the problem again and again in different ways trying to break it down into understandable pieces.
During the development of ChangeSafe I found myself "painted into a corner" on more than one occasion. The code had evolved to a point where it was clear that some core abstractions were the foundation of solving the problem, and the implementation of these core abstractions were at the very center of the solution. Sometimes we were wrong.
Sometimes a core, fundamental abstraction is best understood as a combination or special case of even more fundamental concepts that aren't obvious until later (or much later). This happens fairly often, so you re-implement the relevant code. Often the new abstractions are not completely compatible with the old ones, so you write some scaffolding code or an adapter API, or you bite the bullet and walk through your entire code base and fix up each and every call that used to do things the old way, and make it work the new way.
Sometimes, though, there is simply no way massage the old solution into the new one. You are faced with a tough choice: continue using the old code that doesn't quite do what you want and is becoming increasingly difficult to maintain, discard the entire body of code and start from scratch, or attempt to paper over the problem with some horrific kludge that you say will be temporary (but you know what that means).
Unfortunately, it is hard to tell if you will run into a roadblock like this until you are well down the path. Your new abstraction is making everything so much easier until you realize that in some obscure corner case it is fundamentally incompatible with the old code. Now you have a monstrous pile of code and half of it does things the new way, the other half does it the old way, and the two halves aren't on speaking terms anymore. That's bad.
Now consider that you have a persistent store full of objects that were created and curated by the old code. You cannot delete the old code if you want to access the old objects, but the new code cannot deal with the legacy objects without calling or incorporating the old code. Of course the old code will be confused by the new objects.
On one occasion I ran into a problem of this nature. A certain kind of object needed to be built using some new abstractions, but there were objects in the database that were incompatible with the new world order. I thought of different ways to deal with this. One possible solution was to create yet another abstraction that acted like a bridge between the two models. We'd migrate everything from the legacy system to the bridge system, then once there were no legacy objects, we'd replace the legacy system with the new system and then migrate everything back. "But Bullwinkle, that trick never works!" I can confirm that it does not.
But... in the particular case in point, there was a way it could work if only there were an extra level of indirection. If we could just juggle the underlying physical slot names of the legacy objects without changing the code that manipulates the logical fields, we could fake up a slot that doesn't actually exist in the legacy objects, but does exist in the new ones.
As it turns out, this is easy to do in CLOS by simply adding a couple of methods to the accessors to handle the legacy case. I was delighted to discover this clever escape hatch. It perfectly solved a problem that was starting to look like a depressing amount of very hard work. It almost seemed as if it were designed with this exact problem in mind. And of course, it was. Suddenly, I was very impressed.