I remember as a teenager trying to understand what made rally cars go bang-bang, by modifying the ignition timing among other things, those gases hit a hot exhaust manifold where they create a mini explosion that keeps turbo spooled.
I was working on a 3D CAD software called OpenSCAD, there is this quirk I hated, which is when it's rendering all UI stops responding, there is no cancel button too. So I asked it to "implement the renderer in a separate thread with feedback and the ability to cancel".
It's written in C++ / QT, and can't remember the threading model, I knew Codex ~ version 4 and generally coding agents were excellent at Web type stuff - so I didn't expect much, but it did an alright job.
It was able to build loops of write/compile/verify loops - in the end though, the app would crash, with more work I could probably ask it to debug what was happening.
hell to the no, in between coding sessions, I go out on plenty of sidebars about random topics that help me, the prompter understand the problem more. Prompts in this way are entirely related to context (pre-knowledge) that is not available to the LLMs.