Clickbait content typically uses sensationalized language to entice curiosity and stimulate user interaction, often at the expense of user satisfaction. In this paper, we introduce a spoiler generation system designed to neutralize clickbait by disclosing important information in a clear, informative format. Our system combines three primary components: (1) a RoBERTa-trained spoiler type classification model to determine whether the spoiler must be a phrase, passage, or multi-part; (2) a sequence-to-sequence T5-based model fine-tuned to take the predicted spoiler type as input to produce context-dependent spoilers; and (3) a post-hoc ensemble mechanism that combines predictions from multiple random seeds using edit-distance minimization to improve output consistency. Experimental results on the Webis-Clickbait-22 dataset show that our ensemblemethod substantially outperforms single-model baselines, especially for phrase and multi-part spoilers. These results emphasize the benefits of combining ensembling with spoiler-type conditioning for reliable spoiler generation.