Writing Lisp is AI Resistant and I'm Sad
7 points by spc476
7 points by spc476
I'm having the opposite experience. Claude's been writing a fair amount of Common Lisp for me and after some initial hiccups doing a pretty good job. You can teach it how to do lisp your way with CLAUDE.md files and periodically "reflecting" to save memories. Funny enough I'm also doing some RSS stuff.
A lot of folks are dropping LLM-assisted lisp libraries on the various subreddits, which seems to be causing a lot of hand wringing.
Check out the existing MCPs for the Common Lisp Repl, and libraries for writing MCPs in Common Lisp, some examples:
I think trying to integrate specifically with REPL as implemented for humans (not even an eval-server!) is going too far.
But when I wanted a painfully boring part of some Lisp code generated, local Qwen one-shotted a close-enough draft for me to fix. Real programmers can write Fortran code in any language, and LLM were trained on some of this — pretty sure that code making best use of details of CLOS or syntax-rules could be harder for LLMs, though.
using the REPL as the only interface to Lisp makes the comparison to Python sort of apples and oranges: i suspect OP would have had much more reasonable results having the model make Lisp source files and reading or loading those in the REPL
IMO LLMs have no trouble producing purely functional code, but Lisp can use mutation, macros, effects, user interactivity .. and that becomes much harder to simulate.