Muddled through for a decade with a vague sense that nothing worked as I expected it to. Gradually realised there are many variations with different behaviour,

Muddled through for a decade with a vague sense that nothing worked as I expected it to. Gradually realised there are many variations with different behaviour, obfuscated by dreadful documentation and emergent behaviour in the implementations. Then wrote a regex engine. It's now extremely obvious how regular expressions work as they're very simple. The spurious divergence in syntax and semantics is still infuriating but at least I know what they're supposed to be desugaring to now. Recommended as a worthwhile exercise. Regular expressions have "generate a letter of the alphabet" as a primitive. It might be ascii and use 'b' to generate that letter, or a hex escape like \x42 or similar. The notation varies a bit. Another primitive is "generate the empty string". Then there are compound operations. One regex or another, one followed by another. Intersection, complement. All the set operations, for the reason that a regular expression is literally a notation for a set of strings. Some things like "lookahead" are notation for intersection. The match previous construct, \2 or similar, takes you out of regular expressions but works like checking equality on the fly. Finally anchors, $ or ^ etc, are specific to the match problem. It's still find an element of the denoted set but with some extra constraints on where the element can occur. I'm pretty sure that's all of it. How anchors interact with the set description is a nuisance but seems well formed - I haven't bothered to work through that part yet because I'm mostly interested in string generation, not matching.