An arXiv paper introduces USDCraft, a framework in which a pretrained LLM writes and revises executable programs to turn partial or corrupted meshes into simulation-ready articulated USD assets, without task-specific training. It uses source geometry analysis, which converts the source mesh into a metric textual description separating observed surface from unknown space, plus iterative geometric rechecking that re-encodes each candidate so discrepancies point to program edits. The authors report leading articulation recovery on two benchmarks and validate the approach in real-to-sim-to-real robot manipulation, with assets loading into Isaac Sim without manual adjustment.