Intent Tagging: Exploring Micro-Prompting Interactions for Supporting Granular Human-GenAI Co-Creation Workflows

Research Background and Problems

• Identified Problems or Challenges: This paper highlights the following challenges faced by current Generative Artificial Intelligence (GenAI) systems in content creation:
  1. Alignment of Goals and Content: The generated content often fails to accurately reflect the user's intent.
  2. Prompt Construction Dilemma: Users are often uncertain about how to create effective prompts and understand the AI system's capabilities, leading to trial-and-error interactions.
  3. Lack of Flexibility: Current GenAI systems often force users to adopt predefined workflows, making it difficult to support diverse and non-linear content creation processes.
  4. Uncertainty in the Creation Process: Content creation is iterative and reflective, where users typically cannot clearly articulate their needs and must gradually clarify their intentions.
• Importance of the Research: Addressing these issues can significantly enhance the practicality of generative AI systems in content creation, improving the efficiency and quality of human-AI collaboration.
• Research Motivation and Related Work: Through a literature review, the authors analyzed existing methods for guiding user intent, including graphical user interfaces (GUIs), text, and dialogue. However, these methods involve trade-offs between interactivity and flexibility. This prompted the authors to explore new interaction methods to meet the needs of iterative and diverse workflows.

Solution

• Proposed Method: The authors propose a new interaction paradigm called "Intent Tagging," which supports more granular and non-linear AI collaborative workflows through micro-prompts. Based on this method, they developed a system named "IntentTagger" to enable users to create and modify slide presentations using intent tags.
• Innovations:
  1. Introducing "intent tags" as micro-prompt units, where each tag represents a specific aspect of user intent.
  2. Developing a dynamic tag suggestion mechanism: generating additional suggestions from existing user information, supporting various input forms (e.g., keywords, existing text, images).
  3. Supporting multi-level, non-linear workflows: allowing users to simultaneously operate on global (slide groups) and local (individual slides) generation and editing tasks.
  4. Reducing cognitive load and facilitating intent expression through clearly visualized tag groups and real-time preview functionality.
• Implementation Steps and Key Technologies:
  1. Utilizing the GPT foundational model to generate various suggestions and content.
  2. Designing a graphical user interface based on React.js to present and interact with tags.
  3. Building intent tag suggestion and real-time preview functionalities to support flexibility and control.

Research Outcomes

• Specific Results:
  1. User studies revealed that participants generally found intent tags easier to use compared to existing dialog-based or design gallery-based GenAI systems, offering greater control and higher satisfaction with the results.
  2. The system significantly alleviated the "blank page syndrome," helping users to start more quickly and organize their ideas.
  3. Intent tags demonstrated significant advantages in facilitating human-AI collaboration and helping users clarify their intentions ("meta-intent guidance").
• Advantages Over Existing Solutions:
  1. Compared to chat-based interactions, intent tags provide more precise micro-prompt control and reduce trial-and-error costs.
  2. Users can quickly switch between generation and editing modes in a non-linear manner, greatly improving workflow efficiency.
  3. The system's suggestion functionality offers users diverse creative directions (e.g., additional images, tags), expanding the creative design space.
• Experimental or Evaluation Results:
  1. Compared to current commercial AI-assisted tools (e.g., PowerPoint Copilot), intent tags showed significant improvements in efficiency, task completion rates, and user satisfaction.
  2. Provided researchers with data on user interaction behaviors, including patterns and priorities in tag usage.
• Limitations and Future Directions:
  1. Users initially experienced confusion regarding tag naming and categorization, necessitating further optimization of tag organization and user guidance.
  2. The system currently has limited control over visual styles, and future research should explore how to support locking design elements.
  3. The interaction paradigm needs to be extended to other applications, such as blog creation, video production, and 3D scene design.
  4. Further exploration is recommended on effective integrations of manual editing with AI generation.

Conclusion

This paper introduces intent tags and their interaction system (IntentTagger), showcasing new possibilities for generative AI in supporting creative workflows and validating its effectiveness in enhancing user experience and task efficiency. This not only provides guidance for designing AI-driven content creation tools but also offers a more granular perspective on how to better support "human-AI collaborative creation."