"I Just Need GPT to Refine My Prompts”: Rethinking Onboarding and Help-Seeking with Generative 3D Modelling Tools

Paper Title

"I Just Need GPT to Refine My Prompts”: Rethinking Onboarding and Help-Seeking with Generative 3D Modelling Tools

Publication Info

• Topic area: Onboarding and help-seeking behaviors in generative AI-assisted 3D modeling tools.
• Keywords: Generative AI, 3D modeling, onboarding, help-seeking, prompt engineering, AI chaining, professional workflows, casual users, credit-based systems, HCI.

Background and Problem

• Problem / challenge: Traditional onboarding methods (e.g., tutorials, documentation) are insufficient for generative AI tools, as users must navigate unpredictable outputs, ambiguous prompts, and limited feedback. There is limited understanding of how generative AI reshapes onboarding and help-seeking in creative domains like 3D modeling.
• Significance: Generative AI tools promise to lower barriers for casual users and accelerate workflows for professionals, but new frictions arise, such as the need for effective prompt formulation and interpreting stochastic outputs. Understanding these dynamics is critical for designing systems that support diverse expertise levels.
• Motivation and related work: Prior HCI research has explored learnability barriers in complex software, help-seeking in online communities, and usability challenges in intelligent systems. However, these studies do not address how generative AI changes onboarding and help-seeking behaviors, particularly in creative workflows like 3D modeling.

Solution

• Proposed approach: The study investigates onboarding and help-seeking behaviors in generative 3D modeling tools through an observational study with 26 participants (14 casual users, 12 professionals) using Meshy AI and Spline AI.
• Novelty:
  1. Identification of "prompt-first" mental models and AI chaining (using external AI like ChatGPT for prompt refinement).
  2. Empirical insights into how generative AI reshapes onboarding and help-seeking across expertise levels.
  3. Analysis of credit-based constraints on exploration and iteration behaviors.
  4. Design implications for dual-mode workflows, micro-scaffolds, and credit transparency.
• Procedure and key techniques:
  • Observational study with task-based modeling (text-to-3D, image-to-3D, open-ended design).
  • Semi-structured interviews to capture reflective insights.
  • Analysis of help-seeking behaviors, prompt iteration strategies, and workflow integration.

Results

• Concrete findings:
  • Casual users relied on external AI tools (e.g., ChatGPT) for prompt refinement, while professionals used self-formulated, detailed prompts.
  • Professionals judged AI outputs as not production-ready, while casual users often accepted outputs as "good enough."
  • Credit constraints reduced exploration, with credit-aware participants averaging 2.3 iterations per task compared to 4.4 for credit-unaware participants.
  • Help-seeking was sparse; participants avoided built-in tutorials and documentation, favoring trial-and-error or external resources.
• Advantage over baselines: The study highlights new onboarding and help-seeking behaviors specific to generative AI tools, such as AI chaining and prompt-level iteration, which are not addressed in traditional HCI onboarding frameworks.
• Experiments / evaluation:
  • Participants (n=26) completed three modeling tasks using Meshy AI and Spline AI.
  • Data collection included screen recordings, think-aloud protocols, and post-task interviews.
  • Analysis revealed differences in prompt styles, iteration strategies, and workflow integration between casual and professional users.
• Limitations and future work:
  • Limited to two generative 3D modeling tools and a specific participant demographic.
  • Fixed task order may have influenced behaviors.
  • Future work should explore broader user demographics, additional tools, and systematic variations in credit availability.

Summary

This study examines how generative AI reshapes onboarding and help-seeking in 3D modeling tools, revealing distinct behaviors between casual and professional users. Casual users often relied on external AI tools like ChatGPT for prompt refinement, while professionals used detailed, self-formulated prompts but found outputs unsuitable for production. Credit constraints significantly influenced iteration behaviors, reducing exploration. The findings suggest design implications such as dual-mode workflows, micro-scaffolds for prompt refinement, and credit transparency. These insights contribute to HCI by highlighting how generative AI transforms creative workflows and help-seeking practices.