ABSTRACT Large language models can translate natural language into robot code but often produce physically infeasible behaviours, particularly in tool‐use tasks requiring precise contact reasoning. We propose a closed‐loop framework that integrates code generation, physical simulation, and iterative correction to ensure physically valid tool‐use primitives. Failures such as collisions and misalignments are detected via simulation feedback and corrected by targeted parameter updates while preserving task structure. Across five tool‐use tasks in simulation and real‐world tests, our method achieves a 92% average success rate, representing a 119% improvement over prompt‐based generation and a 16% improvement over vision‐language model‐only correction. These results highlight the importance of simulation‐guided reasoning for robust language‐driven manipulation.