Code as data: LLMs turn out to handle Lisp-based live app editing surprisingly well

Free_Occasion_8354 · reddit · 2026-11-08

A companion post on AffrescoUI focusing on a surprise finding: LLMs handle s-expressions really well.

Same setup: apps are programs in a custom Lisp run by a sandboxed C++/OpenGL engine; model edits are previewed live and can never be self-approved; a two-way MCP server lets agents read and edit running apps.
Core insight: because code is data, each edit is a small, targeted change to a tree, and hooks/advice let the model wrap existing behaviour instead of rewriting it — the author expected models to struggle with Lisp.
Discussion points: how far LLM competence in niche languages extends, keeping model edits small and reviewable, and what an MCP server like this should expose.

More from coding & agent