Large Language Models (LLMs) are increasingly used to support effective and constructive human-to-human communication. A key scenario is helping users express difficult intents, such as criticism and boundary setting without harmful language. Existing safety refusal and detoxification rewriting methods either block harmful responses or inadequately preserve the user’s communicative intent and speaker style. Therefore, we propose constructive communication rewriting, which aims to rewrite harmful utterances in dialogue into safer and more constructive language while preserving user intent and style. To instantiate this task, we construct a large-scale dataset of realistic multi-user conversations by modeling both user persona and harm-inducing contexts. We further use Nonviolent Communication-inspired strategies to guide the rewriting process. Building on this dataset, we develop CoRe, a small, specialized LLM fine-tuned through a two-stage distillation process. CoRe achieves substantially higher harmlessness than strong general-purpose LLMs while maintaining competitive faithfulness and remaining efficient for real-time communication. Our project is released at the following link[<https://github.com/wsfwww/Beyond-Refusal-Constructive-Communication-Rewriting-for-Harmful-User-Language>.].