Maybe we need an LLM interface that is less chatty and unpredictable and focuses instead on boring, predictable execution of commands.

I know, weird idea: what if an LLM just shut up and did what we asked in as predictable a way possible?

Working with an LLM often feels like herding cats.