The rise in online trolling, marked by intentionally posting offensive messages to provoke or disrupt, poses significant challenges to digital interactions. Trolling negatively impacts psychological well-being and disrupts online communities. Our research focuses on AI-based moderation, moving beyond traditional word censorship to accommodate user diversity and mental states in content moderation. We incorporated the discover, define, develop, and deliver stages to create a user-centered comments moderation service. This service was informed and validated through user feedback. Our research included conducting user surveys (n = 106) and in-depth interviews (IDI, n = 9), revealing a diversity in preferences for comment moderation in existing services. Notably, many users expressed a preference for replacing offensive words with alternatives rather than entirely blocking comments. In response to these findings, we designed and developed a stepwise moderation service using current large language models, such as ChatGPT. This service empowers users to choose their preferred level of comment moderation, striking a balance between free expression and a respectful online environment. We tested a working prototype with potential users (n = 9) to evaluate its effectiveness. The results highlight the wide range of user needs and preferences in comment moderation. While there was positive feedback towards progressive comment moderation, it was accompanied by concerns about potential over-moderation and the need to preserve genuine communication. These findings underscore the critical need for nuanced moderation approaches that respect user diversity.