I focus on the semantics first: I try to capture the essence of what I really mean to "say". Then I try to imagine the most idealized language that would let me specify as exactly and easily as possible what I really mean, unambiguously.
If I have access to such a language, I use it. If I don't, but know how to implement it, I might do just that if I think it's worth it in terms of time and effort. Otherwise, I try to think of the next best language that could represent what I mean and that I can use or implement more easily. At some point I get to "really low-level" abstractions like Common Lisp objects, closures and macros...
It's a surprisingly high-quality and slow way to program. I figure at some point I'll have accreted so much language implementation knowledge and infrastructure that I'll be unstoppable because I'll almost always be able to make a good approximation of the best tool for the job quickly.
If I have access to such a language, I use it. If I don't, but know how to implement it, I might do just that if I think it's worth it in terms of time and effort. Otherwise, I try to think of the next best language that could represent what I mean and that I can use or implement more easily. At some point I get to "really low-level" abstractions like Common Lisp objects, closures and macros...
It's a surprisingly high-quality and slow way to program. I figure at some point I'll have accreted so much language implementation knowledge and infrastructure that I'll be unstoppable because I'll almost always be able to make a good approximation of the best tool for the job quickly.