I think your post pretty well illustrates how LLMs can and can't work. Favoriting this so I can point people to it in the future. I see so many extreme opinions on it like from how LLM is basically AGI to how it's "total garbage" but this is a good, balanced - and concise! - overview.
One approach would be to write code first, then run it by AI to get a critique. I think that strikes a good balance between avoiding atrophy and still getting the benefits of the tool.
If I create an executable with SBCL's save-lisp-and-die then run that executable, it presents a repl. Not really a development scenario, but more of a "running application" scenario. But, even then, if the executable creates a swank server, I can connect to that from SLIME.
I like LoL alright too. I was only maybe a year or so into CL when I read it, so I didn't really notice a lot of what it's criticized for. On the one hand, I'm a little sad that I see it bagged on so much (that link from beach's site, IRC, reddit) since I felt I got decent enough value out of it, but OTOH it's good and healthy to point out problems.
The abstractions used in Calendrical Calculations sound good - and echo what I've seen elsewhere - so, based on your comment I'm now more likely to read it, so thank you for that.
My opinion of LOOP started to change when I read (the much maligned) "Land of Lisp" and went over that "periodic" diagram in TFA. Seeing the elements of LOOP broken down like that went a long way to get me to overcome my original aversion to it.
edit SICP has examples on how to implement streaming (in Scheme).