I have so far not been able to get a major LLM to generate fully functional Typst code, no matter how much context I try to put into it. The models do not seem to currently understand Typst's concept of modes (code, markup and math) and especially in code mode suffers from heavy hallucination of syntax and/or semantics.