Comment by overspeed

Did something similar with a previous incarnation of OpenAI's LLMs - codex.

Far more constrained no doubt but it made some things convenient like finding obscure `ls` flags.