There are only so many language implementers, and many opportunities to innovate in language design become obvious when it's your turn to decide what to implement.
For example, I made macros and axioms first-class citizens of the language as special cases of the function type, because using an interpreter allowed me to decide how things should be evaluated at runtime. I accidentally found a trivial way to bootstrap (quote) in my language. I also found a way to make a much more general reader macro system by simply writing a modular value reader that can be extended by adding new elements to a list of parsers. This make non-lisp syntaxes like JSON much easier to implement than with the CL reader macro system where reader macros are bound to some special parameters.
Note for context: In my case, the Lisp interpreter is under 3KLoC, and most of the abstractions are bootstrapped from macros and reflection primitives, or primitive that are shadowed by higher level builtins.