The topic of my research right now is a subset of this; it essentially researches the quality of the outputs of LLMs, when they're writing tight-fitting DSL code, for very context-specific areas of knowledge.

One example could be a low-level programming language for a given PLC manufacturer, where the prompt comes from a context-aware domain expert, and the LLM is able to output proper DSL code for that PLC. Think of "make sure this motor spins at 300rpm while this other task takes place"-type prompts.

The LLM essentially needs to juggle between understanding those highly-contextual clues, and writing DSL code that very tightly fits the DSL definition.

We're still years away from this being thoroughly reliable for all contexts, but it's interesting research nonetheless. Happy to see that someone also agrees with my sentiment ;-)