Comment by zos_kia

I've found this approach brings slightly better result indeed. Let the model "think" in natural language, then translate it's conclusions to Json. (Vibe checked, not benchmarked)