Large Language Models (LLMs) have shown promise in generating formal representations such as PDDL in Classical Planning, capable of producing parsable and solvable code; however, despite these recent breakthroughs, they are limited to the natural language ambiguity of the user’s descriptions of the model and can result in semantically incorrect or infeasible real-world plans. We propose FixMyPlan, a general framework that leverages the common sense capabilities of LLMs to judge the semantics of error-prone PDDL plans and back prompt to fix their corresponding models in a closed-loop fashion, produce coherent plans—all while minimizing human intervention. We conduct experiments on 5 f lawed PDDL domains, producing solvable—yet incorrect plans: Blocksworld, Logistics, Mystery Blocksworld and Logistics, and our self-produced domain. We aim to analyze common pitfalls such as semantic ambiguity, unintentional constraints, and logical inconsistencies that hinder effective plan generation and alignment with real-world tasks.