Interaction Patterns in Human–LLM Collaborative Writing Tools

Abstract

Large language models are increasingly embedded in professional writing and documentation platforms. Despite growing adoption, actual interaction behavior between users and language models remains insufficiently understood. This paper examines usage patterns from a commercial writing assistant deployed across administrative, technical, and customer support workflows. Analysis focuses on prompt reformulation, revision cycles, and task completion strategies. Most users engaged in iterative refinement rather than relying on single-response outputs, and explicit control over tone and structure was frequently requested. Interface prototypes incorporating guided templates and contextual cues were evaluated in pilot studies. Observations indicate that transparency and controllability influence long-term acceptance more strongly than raw generation quality. Effective collaboration depends on interface design as much as on model capability.