DuoDrama: Supporting Screenplay Refinement Through LLM-Assisted Human Reflection

Paper Title

DuoDrama: Supporting Screenplay Refinement Through LLM-Assisted Human Reflection

Publication Info

• Topic area: AI-assisted tools for screenplay refinement and human reflection.
• Keywords: AI-assisted reflection, screenplay refinement, performance theory, multi-agent systems, creative writing, LLM, human-AI collaboration, feedback generation, ExReflect, DuoDrama.

Background and Problem

• Problem / challenge: Existing AI tools for screenwriting fail to coordinate internal (character-focused) and external (story-wide) perspectives during screenplay refinement, resulting in fragmented or misaligned feedback.
• Significance: Coordinating these perspectives is crucial for effective screenplay refinement, which shapes the final narrative by addressing scene logic, character motivation, and thematic coherence.
• Motivation and related work: Prior AI systems have used role-playing to provide feedback but typically focus on either internal or external perspectives in isolation. Multi-agent systems offer diverse perspectives but lack coherence and alignment, leaving users to reconcile fragmented feedback. This paper addresses these gaps by integrating both perspectives into a unified feedback system.

Solution

• Proposed approach: DuoDrama, an AI system powered by the Experience-Grounded Feedback Generation Workflow for Human Reflection (ExReflect), which coordinates internal and external perspectives to support screenwriters’ reflection.
• Novelty:
  1. Introduction of ExReflect, a workflow inspired by performance theories (Stanislavski’s immersive embodiment and Brecht’s critical distancing).
  2. Integration of ExReflect into a multi-agent architecture to handle multiple characters in screenplays.
  3. Demonstration of DuoDrama’s effectiveness in improving feedback quality, alignment, and screenwriters’ reflection through a user study.
• Procedure and key techniques:
  1. ExReflect Workflow: Sequentially adopts two roles—an experience role (character’s internal perspective) and an evaluation role (actor’s external critique)—to generate feedback grounded in personal experience.
  2. System Design: Multi-agent architecture assigns one agent per character, with each agent generating inner thoughts and feedback based on evolving screenplay context.
  3. Feedback Types: Instant feedback (line-specific) and post-hoc feedback (scene-level) are generated and evaluated for quality, alignment, and timing.
  4. User Interface: Four-panel interface for uploading screenplays, enacting roles, reviewing feedback, and marking valuable insights.

Results

• Concrete findings:
  • DuoDrama achieved an average SUS score of 84.46, indicating high usability.
  • Significant improvements in feedback quality (e.g., content richness, specificity) and alignment (e.g., character motivation, plot pacing) compared to baseline conditions.
  • Enhanced perceived effectiveness, depth, and richness of reflection, with participants reporting increased revision motivation and insight.
• Advantage over baselines:
  • Outperformed Exp-PE (experience role only), Eval-NoPE (evaluation role without personal experience), and Rev-NoPE (industry-standard reviewer) in multiple dimensions, including emotional insight, alignment, and reflection richness.
  • Feedback grounded in personal experience was more actionable and contextually aligned than generic or detached critiques.
• Experiments / evaluation:
  • Two-session user study with 14 professional screenwriters.
  • Session 1: Participants used DuoDrama on their screenplays and rated usability and feedback quality.
  • Session 2: Comparative evaluation of DuoDrama against three baseline conditions using Likert-scale ratings and qualitative interviews.
• Limitations and future work:
  • Limited to text-based screenwriting; future work could explore multi-modal grounding (e.g., visual or auditory cues).
  • Short-term evaluation; longer-term studies are needed to assess sustained impact.
  • Adaptive mechanisms for feedback timing and quantity could improve personalization.

Summary

DuoDrama introduces a novel AI system for screenplay refinement that integrates internal and external perspectives using the ExReflect workflow. By grounding feedback in personal experience and providing critical evaluation, DuoDrama improves feedback quality, alignment, and screenwriters’ reflection. A user study with 14 professionals demonstrated its advantages over existing approaches, highlighting its ability to surface overlooked issues and motivate refinement. Future directions include expanding to multi-modal feedback, adaptive personalization, and broader applications in human–AI collaboration.