Abstract: To address the challenge of balancing minimal rewriting and causal consistency in counterfactual story generation, this paper proposes a three-stage counterfactual story generation method based on structure decoupling and semantic guidance. Firstly, an explicit causal skeleton decoupling module is constructed at the structural level. The U/V orthogonal decoupling mechanism is adopted to separate invariant background information and causal variation information from the original ending, and the implicit semantic separation results are further mapped into explicit, interpretable causal skeletons at the text level. Secondly, a semantic-guided learning module is designed at the semantic level. Contrastive learning is utilized to map premises and counterfactual conditions into the target causal semantic space to obtain semantic guidance vectors. Finally, a skeleton-semantic fusion controllable generation module is built at the generation level, which unifies explicit skeleton constraints and target semantic guidance into the counterfactual ending generation process. Experimental results on the TIMETRAVEL dataset demonstrate that the proposed method achieves overall performance superior to multiple baseline models. It yields prominent performance on the core causal consistency metric ENTScore and the comprehensive metric HMean. While maintaining the original narrative storyline and minimal necessary rewriting, the method significantly improves the causal-logical consistency between generated outputs and counterfactual conditions.