Paper deep dive
"I Just Need GPT to Refine My Prompts": Rethinking Onboarding and Help-Seeking with Generative 3D Modeling Tools
Kanak Gautam, Poorvi Bhatia, Parmit K. Chilana
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/1/2026, 1:30:26 AM
Summary
This study investigates how 26 participants (14 casual, 12 professional) use generative AI tools for 3D modeling. It finds that users bypass traditional tutorials in favor of a 'prompt-first' approach, often using external LLMs like ChatGPT for recursive support. The research highlights that while generative AI lowers entry barriers for casual users, it introduces new frictions for professionals who struggle with the lack of precision and production-readiness in AI-generated outputs.
Entities (5)
Relation Signals (3)
Kanak Gautam â authored â I Just Need GPT to Refine My Prompts
confidence 100% ¡ Kanak Gautam, Poorvi Bhatia, and Parmit K Chilana. 2026. âI Just Need GPT to Refine My Promptsâ
Generative AI Tools â replaces â Traditional Tutorials
confidence 90% ¡ participants skipped tutorials and manuals, relying on trial and error... the prompt box became the entry point for learning
Casual Users â used â ChatGPT
confidence 90% ¡ some casual users turned to external LLMs for prompts... many preferring to go to ChatGPT instead.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Learning to use feature-rich software is a persistent challenge, but generative AI tools promise to lower this barrier by replacing complex navigation with natural language prompts. We investigated how people approach prompt-based tools for 3D modeling in an observational study with 26 participants (14 casuals, 12 professionals). Consistent with earlier work, participants skipped tutorials and manuals, relying on trial and error. What differed in the generative AI context was how and why they sought support: the prompt box became the entry point for learning, collapsing onboarding into immediate action, while some casual users turned to external LLMs for prompts. Professionals used 3D expertise to refine iterations and critically evaluated outputs, often discarding models that did not meet their standards, whereas casual users settled for "good enough." We contribute empirical insights into how generative AI reshapes help-seeking, highlighting new practices of onboarding, recursive AI-for-AI support, and shifting expertise in interpreting outputs.
Tags
Links
- Source: https://arxiv.org/abs/2603.29118v1
- Canonical: https://arxiv.org/abs/2603.29118v1
Trouble viewing inline? Open PDF directly â
Full Text
100,790 characters extracted from source content.
Expand or collapse full text
âI Just Need GPT to Refine My Promptsâ: Rethinking Onboarding and Help-Seeking with Generative 3D Modeling Tools Kanak Gautam Computing Science, Simon Fraser University Burnaby, British Columbia, Canada kga74@sfu.ca Poorvi Bhatia Computing Science, Simon Fraser University Burnaby, British Columbia, Canada poorvi_bhatia@sfu.ca Parmit K Chilana Computing Science, Simon Fraser University Burnaby, British Columbia, Canada pchilana@sfu.ca Abstract Learning to use feature-rich software is a persistent challenge, but generative AI tools promise to lower this barrier by replacing com- plex navigation with natural language prompts. We investigated how people approach prompt-based tools for 3D modeling in an ob- servational study with 26 participants (14 casuals, 12 professionals). Consistent with earlier work, participants skipped tutorials and manuals, relying on trial and error. What differed in the generative AI context was where and how they sought support: the prompt box became the entry point for learning, collapsing onboarding into immediate action, while some casual users turned to external LLMs for prompts. Professionals used 3D expertise to refine iterations and critically evaluated outputs, often discarding models that did not meet their standards, whereas casual users settled for âgood enough.â We contribute empirical insights into how generative AI reshapes help-seeking, highlighting new practices of onboarding, recursive AI-for-AI support, and shifting expertise in interpreting outputs. CCS Concepts ⢠Human-centered computingâHuman computer interac- tion (HCI). Keywords Generative 3D Modeling, Help-Seeking, Software Learnability, Men- tal Models ACM Reference Format: Kanak Gautam, Poorvi Bhatia, and Parmit K Chilana. 2026. âI Just Need GPT to Refine My Promptsâ: Rethinking Onboarding and Help-Seeking with Generative 3D Modeling Tools. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI â26), April 13â17, 2026, Barcelona, Spain. ACM, New York, NY, USA, 16 pages. https://doi.org/10.1145/3772318. 3791124 1 Introduction Generative AI (Gen AI) tools promise to expand the reach of feature- rich design tools. By producing images, text, or 3D models from natural language prompts, they make it possible to start creating without first navigating the steep learning curves of feature-rich software. But insights from HCI remind us that lowering entry This work is licensed under a Creative Commons Attribution 4.0 International License. CHI â26, Barcelona, Spain Š 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2278-3/2026/04 https://doi.org/10.1145/3772318.3791124 barriers often creates new forms of friction: unfamiliar affordances, opaque errors, and invisible labor such as troubleshooting, prompt iteration, or consulting external resources [22,45,62]. How, then, do users actually onboard these generative design systems, and where do they seek help when things go wrong? We examine this question through the case of 3D modeling, a domain long known for steep learning curves and technical bar- riers. Advances in large multimodal models now allow users to produce 3D models from text-based prompts or sketches, promis- ing faster workflows, richer exploration, and lower entry barriers [42,51,53,54]. These possibilities matter for two distinct groups of creators: professionals, defined here as individuals with estab- lished expertise in 3D modeling through industrial practice, formal education, or freelance production; and casual makers, who are individuals with little or no prior 3D experience but are motivated to explore 3D design and experiment with walk-up-and-use 3D printing services that allow first-time users to begin meaningful ac- tivity without prior preparation or training [39]. For professionals, generative systems can accelerate repetitive tasks, spark new direc- tions in ideation, and complement production workflows [54,80]. For casuals, such systems can open creative avenues that were pre- viously difficult to learn, allowing them to bypass years of technical training and produce tangible 3D artifacts [31]. However, despite this promise, there is limited understanding of how generative AI changes the actual work of onboarding and learn- ing new software. In this paper, we use onboarding to refer to how users begin working with a new software or digital environment, how they learn its capabilities, explore its features, and establish an initial workflow [3,16]. Prior HCI research has documented learn- ability barriers in complex software [23,25,47,82], help-seeking in online communities [37,38], and usability breakdowns in in- telligent systems [46]. Yet, these perspectives offer limited insight into how generative AI reshapes the cognitive and interactional work users must perform [15]. In creative, spatial workflows like 3D modeling, interactions are shaped by unpredictable outputs, ambiguous prompt affordances, and limited feedback when sys- tem interpretations fail. These challenges raise questions about how users, especially those with different levels of expertise navi- gate breakdowns, refine inputs, and decide when to seek support. Without attending to these dynamics, we risk overlooking the very frictions that could limit the broader participation in design that generative AI tools are meant to support. Understanding how these frictions differ across expertise levels is crucial for designing AI systems that truly augment, rather than complicate, creative work [15,60]. Motivated by these insights, our research questions were: arXiv:2603.29118v1 [cs.HC] 31 Mar 2026 CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana Figure 1: Example of a Generative 3D modeling tool (Meshy AI) used in the study. Its key features include: (A) Prompt input panel for describing the desired 3D model, including style and generation settings; (B) Built-in AI help option that refines vague prompts into detailed, structured prompts for more accurate generation; (C) Built-in library of example prompts providing ready-to-use templates for diverse modeling tasks; (D) Built-in tutorial link offering step-by-step guidance on using the tool effectively; (E) Credit and upgrade indicators showing remaining credits and subscription options. Spline AI interface was similar to this interface as well. Despite help and examples built into the UI, they were largely ignored by participants, many preferring to go to ChatGPT instead. â˘RQ1: How do casual and professional users form initial men- tal models when interacting with prompt-driven generative 3D modeling tools? â˘RQ2: How do users seek help and refine their workflows as they iterate designs with these generative 3D modeling tools? We conducted an in-depth observational study with 26 partici- pants (14 casuals, 12 professionals) using state-of-the-art prompt- driven 3D modeling systems (Meshy [61] and Spline [76]), where participants completed various modeling tasks using natural lan- guage. We also conducted semi-structured interviews to capture further insights into their experiences and decisions about how and when to seek support. Through task-based observations, think- aloud protocols, and interviews, we analyzed how participants nav- igated onboarding, formulated prompts, sought help, and refined their workflows over multiple iterations, with particular attention to differences between casual and professional users. Our findings reveal new tensions in how generative AI lowers entry barriers for casual makers while simultaneously introducing frictions for professionals whose expectations for precision and workflow integration exceed current capabilities. Thematic analysis revealed three critical insights: both casual and professional users shared a âprompt-firstâ mental model, treating the text box as the primary entry point and bypassing tutorials or documentation; casual users frequently leveraged external AI tools such as ChatGPT to co-construct prompts, echoing concerns about prompt writing as a distinct practice of trial and error [52,59]. Professionals often discarded AI-generated models as ânon-production-readyâ, whereas casual users expressed satisfaction with outputs they perceived as sufficient for visualization or hobbyist 3D printing. Our paper contributes empirical insights into how 26 casual and professional users onboard to generative 3D modeling tools. Our analysis reveals practices such as treating the prompt as a tuto- rial, using recursive âAI-for-AIâ support, and engaging in resource- rational iteration shaped by credit-based systems. We also derive design implications for how generative AI can lower entry barriers for newcomers while supporting the precision and control required by professionals. By situating these insights within broader HCI work on humanâAI collaboration, interpretability, and co-creativity, we extend understandings of onboarding and help-seeking in com- plex software. Our results suggest that in the era of Gen AI, software onboarding is emerging as a language-mediated experimentation Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain rather than traditional interface exploration [14]. We highlight what is new: the emergence of AI-to-AI help-seeking and AI chaining as primary strategies for onboarding and troubleshooting. At the same time, our analysis shows that usersâ underlying motivations remain consistent with long-standing patterns identified in prior HCI and learning work [2,13,14,32,36,44].While our study focuses on 3D modeling, the observed dynamics likely extend to other domains where generative AI is embedded into feature-rich creative tools. 2 Related Work While our study highlights how generative AI reshapes onboarding and help-seeking, these issues are not new to HCI. Prior work has long examined how users learn feature-rich software, seek support through documentation and communities, and adapt to intelligent systems. HCI has grounded these behaviors in theories of learnability, self-regulated learning, distributed cognition, and trust in automation [14,32,44]. What remains less understood is how generative AI changes these dynamics, collapsing tutorials into prompts, shifting help-seeking across tools, and reframing expertise. We therefore turn to related work on software learnability, help- seeking, and generative AI in creative practice. 2.1 Learning and Help-Seeking in Complex Software Feature-rich creative tools, such as CAD systems, game engines, and 3D modeling platforms, have long been recognized for their steep learning curves. Prior HCI studies document the vocabulary prob- lem [30,39,72], where novices struggle to find help using correct terminology, and show that users often bypass manuals or tutorials [67]. This pattern has been classically described as the paradox of the active user [14], which explains why people prioritize immedi- ate task progress over systematic learning even when this slows long-term performance. Documentation, video tutorials, and com- munity forums provide flexible but inconsistent support: they may be outdated, difficult to search, or misaligned with the userâs goals [10,73,85]. Alongside such embedded supports, prior work has shown that trial-and-error remains a dominant learning strategy in complex software, where users iteratively test commands or features until achieving a desired outcome. Studies of CAD, pro- gramming environments, and creative tools consistently observe this incremental process as a practical, if sometimes inefficient, way to onboard and make progress [59,81]. This illustrates how on- boarding and help-seeking have historically been distributed across multiple pathways, each carrying friction. Researchers responded by embedding in-situ support systems [9,20,24] such as interactive walkthroughs, examples, or context-aware agents directly into in- terfaces to scaffold learning within workflows. These systems aim to reduce friction by surfacing assistance within the userâs workflow, lowering barriers to experimentation and learning. However, these approaches often target deterministic software behaviors and may fall short when integrated into generative systems with stochastic outputs, highlighting the need for more adaptable support that flexibly responds to unpredictable model behavior. To the extent these approaches can be integrated into generative systems with stochastic outputs and unpredictable model behavior is an open research question. Our work adds new insights about how onboard- ing and help-seeking unfold in generative AIâassisted 3D modeling, where users must not only learn tool interfaces but also grapple with interpreting unpredictable outputs. We show how both ca- sual and professional users navigate these challenges by blending self-written prompts, external resources, and AI-based guidance, revealing new forms of friction and support-seeking strategies that extend beyond those documented in prior studies of deterministic software. 2.2 Invisible Labor and Hidden Costs in Intelligent Systems Beyond software learning, recent work in HCI and software engi- neering is illuminating the invisible labor that sustains interaction with complex and intelligent systems [15,29,35,57,60,66,83]. Stud- ies of end-user programming and professional software use show that users often rely on trial-and-error loops, incremental debug- ging, and extensive online searches, forms of work that are rarely visible in final output [10,14,47]. With the advent of large language models, this hidden effort shifts toward prompt crafting and itera- tive refinement. Research on AI-assisted coding demonstrates how novices may treat AI assistants as black-box tutors while experts integrate them selectively, both groups expending unseen cognitive and emotional labor to vet outputs, recover from errors, or adapt workflows. Broader humanâAI collaboration studies underscore how unpredictability, lack of transparency, and uneven control ex- acerbate these hidden costs [68,79,84]. This body of work provides a critical lens: while generative AI appears to âlower barriers,â it may instead redistribute labor into subtler, less visible forms that shape trust, efficiency, and learning. Our study complements this body of work by examining invisible labor in the domain of genera- tive 3D modeling. Unlike coding or end-user programming, where trial-and-error loops or debugging are already established practices, 3D modeling introduces distinct forms of hidden effort: crafting prompts that balance form and function, repeatedly iterating when outputs misalign with intent, and turning to external resources when system guidance falls short. By surfacing how both casual and professional users distribute their labor across prompting, ex- ternal help-seeking, and post-processing, our findings highlight how generative systems may not simply reduce complexity but shift it into less visible forms of interaction. In doing so, we extend HCI conversations on invisible labor from software engineering contexts into creative design tools, showing how unpredictability and stochasticity uniquely reshape onboarding and learning in this domain. 2.3 Generative AI and Prompt-Based Creative Practice In addition to software development, recent advances in generative AI have opened new possibilities in creative domains such as art, text, music, and 3D modeling [5,27]. Prompt-based systems like DALL¡E and Firefly illustrate how natural language interfaces col- lapse low-level commands into expressive requests, while research has begun examining prompting as a skill that shapes quality, di- versity, and control of outputs [1,55,71]. In 3D, models such as DreamFusion, Magic3D, and Shap-E demonstrate pipelines that CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana Table 1: Overview of participant demographics, including gender, age group, academic or professional background, and tool expertise of the professionals) across all 26 participants in the study. Participant IDs with a âCâ indicate casual users, while IDs with a âPâ indicate professional users. IDAgeBackground3D Tool ExpertiseIDAgeBackground3D Tool Expertise C0119â24 (M)Computer ScienceNoneC10 19â24 (F)Health SciencesNone C0219â24 (F)DesignerNoneC11 19â24 (F)Computer ScienceNone C0319â24 (F)International StudiesNoneC12 25â34 (F)Web DesignNone C0419â24 (F)StatisticsNoneC13 55â64 (F)PhysicsNone C0519â24 (F)ArtsNoneP14 19â24 (F)Computer ScienceAutoCAD P0625â34 (M)Game DesignBlender, MayaC15 25â34 (M)Computer ScienceNone P0725â34 (M)3D PrintingAutoCADP16 25â34 (F) Interactive Arts & Tech.AutoCAD, Rhino C0819â24 (F)Computer ScienceNoneP17 35â44 (M)Game DesignBlender, Maya P0925â34 (F)3D PrintingAutoCAD, RevitP18 35â44 (M)AR/VR3Ds Max P19 19â24 (GQ)Computer ScienceUnity, MayaC20 25â34 (F)Computer ScienceNone P2135â44 (M)Computer ScienceFusion360, AutoCADC22 19â24 (F)Health ScienceNone P2319â24 (M)Electrical Eng.AutoCADC24 25â34 (M)Computer ScienceNone P2519â24 (F)Computer ScienceFusion360P26 19â24 (M)Computer ScienceFusion360 translate text or images into meshes, while sketch-to-3D systems more closely align with traditional design workflows [17,42,53,70]. These innovations have given rise to casual-facing tools (e.g., Meshy AI, Spline AI) that echo âwalk-up-and-useâ fabrication paradigms, as well as professional use cases where AI outputs are exported for refinement in Blender or Unity [8,12,78]. Yet HCI scholarship re- mains limited in understanding how real users actually adapt: casual users may treat AI as a shortcut to producing printable forms, while professionals critique models against production standards and dis- card outputs deemed unusable. This highlights a gap: while prior work studies prompting and creative AI broadly, little is known about how onboarding and help-seeking play out when generative AI becomes the entry point to complex domains like 3D modeling. Our study addresses this gap by examining how both casual and professional users onboard generative AIâassisted 3D modeling tools and seek help when encountering uncertainty. We show how participants engage in trial-and-error loops, and turn to external resources when system support falls short. By surfacing these prac- tices, we extend HCI scholarship on prompt-based creative practice beyond 2D media and coding, situating it within 3D workflows where precision, production constraints, and usability demands create unique forms of friction. In doing so, we highlight how gen- erative AI does not simply collapse complexity but redistributes it into new learning and labor practices. 3 Method The goal of this research is to explore how users with varying levels of expertise in 3D design, ranging from casual makers with little to no prior experience to professional designers with formal train- ing onboard, seek help when navigating generative AI modeling tools for the first time. We take an interpretivist stance [11], aim- ing to surface participantsâ meaning-making and invisible labor, rather than to measure performance outcomes. We used a quali- tative approach as it is well-suited for exploring how practices of onboarding and help-seeking unfold in situated, real-world use of generative tools. We started with an observational task-based study using two popular generative 3D modeling tools and followed-up with in-depth interviews to capture usersâ reflective understanding of the challenges they faced, the strategies they adopted, and how they would prefer such systems to support creative practices and learning processes. 3.1 Participant Recruitment and Grouping We recruited a total of 26 participants (10M, 15F, 1 Gender Queer), as shown in Table 1 and divided them into two categories: ⢠Casual Users (C): Participants who had little or no formal training in 3D modeling and were novices in this space [39, 40]. This group included students from both technical and non-technical backgrounds, hobbyists interested in creative tools, and other users curious about generative 3D design. None of them had used CAD tools extensively and were not familiar with professional modeling workflows. (n = 14) â˘Professional Users (P): Participants who had formal training and/or industry experience in 3D modeling, such as CAD designers, architects, industrial designers, game designers and 3D artists. All had previously used tools such as Fusion 360, Rhino, Blender, or SolidWorks in professional settings and were familiar with modeling constraints for production and 3D printing. (n = 12) Casual users were recruited through university mailing lists and flyer advertisements, while professionals were recruited through professional design networks, local makerspaces, and snowball sampling. Our goal was to cover a broad spectrum of experience levels in 3D modeling, ranging from absolute novices to trained professionals. We ensured a diverse sample across domains, ed- ucational backgrounds, and age groups. Table 1 summarizes the demographics, backgrounds, and experience levels of our partic- ipants. Most of our participants never used any Generative 3D modeling application before. As a research team with backgrounds in HCI and experience in both professional 3D design and novice creative tool use, we recognize our positionality shaped what we attended to in data col- lection and interpretation. We practiced reflexivity through regular debriefs, challenging our assumptions about expertise, creativity, and tool usability. Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain (a) Target model used in Task 1(b) Target model used in Task 2 Figure 2: Target reference models provided to participants for modeling tasks. (a) Task 1: A glass-top wooden table (adapted from [47]); (b) Task 2: A desk lamp with an adjustable arm. These tasks were designed consulting experts in 3D modeling and were designed in target generative AI tool (Meshy or Spline). 3.2 Study Context and Tool Selection For our study, we sought tools that would allow users to generate 3D assets using natural language prompts, reference images, or texture specifications. After exploring available platforms and consulting with 3D design experts, we selected two widely used web-based generative 3D modeling systems, Meshy AI [61] and Spline AI [76], that represent the current wave of commercially deployed AI-powered modeling tools. A key characteristic of these systems is that they operate on a credit, or usage-based pricing model: each model generation, re- finement, or upscaling action consumes a predefined number of credits. This pattern is not unique to 3D modeling but increasingly common across the broader GenAI ecosystem. Major creative-AI platforms such as Runway (video generation) [74], Adobe Firefly (image and design generation) [1], and Sora [65] also meter us- age through generative credits; similarly, multimodal APIs from OpenAI [64], Stability AI [77] and Google rely [33] on token- or generation-based billing rather than unlimited access. These exam- ples reflect a broader industry shift toward usage-based pricing for AI-enhanced creative work, particularly in feature-rich domains where computational cost varies by output quality and model com- plexity. Although our study did not explicitly aim to investigate credit behavior, credit constraints organically emerged as a strong influence on user decision-making during pilot testing. Hence, to maintain the ecological validity of the study, we decided to test participants under credit-based situations. 3.3 Task Design and Rationale To investigate how users approach and learn generative 3D mod- eling, we designed three tasks reflecting common entry points in creative and professional workflows. These tasks elicited behaviors ranging from initial prompt instincts to visual input interpretation and open-ended ideation. We refined them through pilot testing and consultation with practicing 3D designers. Task 1: Text-to-3D Generation The first task was to generate a 3D model using the toolâs natural language prompt interface. This task examined usersâ initial mental models when approaching generative 3D modelingâwhether they described high-level features (e.g., overall shape or intended func- tion) or low-level ones (e.g., specific parts, materials, or fine details of the model). We also observed strategies for handling uncertainty, such as trial and error and use of help within the application or through external resources (participants were encouraged to use any help resource of their choice for all tasks). The target object was a glass-top table with a wooden base, as shown in Figure 2a, adapted from Kiani et al.âs [47] study. We selected this as it has a moderately complex structure involving mul- tiple materials and surface expectations, useful for assessing how participants translate visual ideas into text prompts. The reference model was created in advance by the researchers using the same AI tools (Meshy and Spline) and refined after consulting professionals in this domain. This task surfaced how participants (especially those unfamiliar with the 3D modeling vocabulary) navigate the prompt formulation, feature interpretation, and model inspection when no explicit visual cues are provided. Task 2: Image-to-3D Generation: Sketch-Based Visual Modeling In Task 2, participants generated a 3D model from a sketch using the generative 3D toolâs image-to-3D feature. This task was inspired by early findings from our pilot study, where multiple professional participants expressed a preference for visual input as a more natural part of their creative process, as many workflows begin with sketches, moodboards, or reference images rather than detailed text prompts. We provided participants with a minimal line drawing of a small side table lamp, as shown in Figure 2b, and asked to generate a 3D model that aligns with the sketch. This allowed us to assess how participants interpret and control the fidelity of the AI output against a visual target, and how they perceive tool capability when working with concrete visual constraints. This task revealed differences between casual and professional participants in leveraging visual scaffolds and evaluating the quality of 2D to 3D translation across generative platforms. CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana Task 3: Open-Ended Generation: Creative or Goal-Oriented Design The third task was tailored by the participant group. Casual users were asked to design a simple, fun, or useful object for 3D printing, revealing their motivations, ideation strategies, and perceptions of printability and tool constraints. Professional users were asked to create a basic model for potential professional use (e.g., a prototype enclosure), allowing us to examine how they assessed mesh quality, editability, and readiness for downstream CAD or production tools. Overall, this task surfaced where participants saw value in AI assistance, whether for ideation, sketching, rapid prototyping, or final production, and how expectations shifted with expertise. 3.4 Study Procedure Each study session was conducted in person in the lab or remotely over Zoom and lasted between 60 and 75 minutes. All remote par- ticipants joined from their personal computers in their own envi- ronments to preserve ecological validity and simulate natural use conditions. The procedure was structured into three phases: orien- tation, task execution with concurrent think-aloud, and a post-task semi-structured interview. 3.4.1 Orientation and Setup (5â10 min). Participants first com- pleted a pre-study questionnaire designed to capture demograph- ics, prior experience with 3D modeling tools, and familiarity with generative 3D design. We next randomly assigned participants to either the Meshy AI or the Spline tool condition using researcher- provided accounts. The researcher running the study provided a brief orientation to the platform interface, covering basic navigation and functional features. Importantly, no task-specific examples or strategies were shared; this decision was intentional to recreate the self-directed onboarding conditions under which users typically encounter generative AI tools. Participants were informed that they could freely use any help built into the tool and any external resource they may normally rely on (e.g., Google, YouTube tutori- als, platform documentation, or ChatGPT) but that the researcher would not provide direct task guidance. 3.4.2 Task Execution and Think-Aloud (35â45 min). Participants completed the three modeling tasks in a fixed order: text-to-3D gen- eration, image-to-3D generation, and the open-ended design task. Tasks were presented in a fixed order to scaffold participants from lower- to higher-constraint activities, reflecting how generative tools are typically encountered. While we considered counterbal- ancing, we prioritized a consistent progression to ease learning for novices and to keep the focus on interaction patterns rather than comparative performance. To explore how credit systems shape behavior, we introduced two contrasting contexts. Half of the participants (n=13) were credit- aware: they were told about the platformâs credit-based generation model and given just enough credits to attempt each task twice. This setting foregrounded how resource limits influenced strategies such as cautious iteration, prompt abandonment, or seeking external assistance. The remaining participants (n=13) were credit-unaware: they received accounts with unrestricted credits and no mention of the system, reflecting the experience of encountering such tools without explicit awareness of usage constraints. This contrast was intended to surface potentially diverse strategies and perceptions, not to produce comparative performance metrics. During task execution, all participants were instructed to think aloud, verbalizing their reasoning, frustrations, and decision-making in real time, including the remote participants who additionally shared their screens. The researcher captured prompt formulations, credit use (or absence of credit concerns), tool navigation patterns, and any external help-seeking behavior through screen record- ings. The researcher remained a silent observer, intervening only in the case of technical difficulties, to preserve the authenticity of participantsâ self-directed learning and troubleshooting. After completing each task, participants filled out a brief 7-point Likert scale questionnaire to rate their perceptions of the platform for that task, their satisfaction with the generated outputs, the predictability and trustworthiness of the systemâs behavior, and their perceived likelihood of using such a workflow in the future. Embedding the questionnaire after each task allowed us to capture participantsâ impressions while they were still fresh, avoiding ret- rospective bias and enabling direct comparison of attitudes across tasks. 3.4.3Post-Task Semi-Structured Interview (15â20 minutes). We briefly interviewed each participant to probe into participantsâ expecta- tions of generative 3D modeling tools, the troubleshooting strate- gies they employed for prompt iteration and error correction, and their evaluation of the usefulness and trustworthiness of AI out- puts. We also asked participants to situate these tools in relation to their broader creative or professional practices. For professional participants, we further explored perceptions of model quality and compatibility with downstream workflows (e.g., CAD tools, 3D printing pipelines). The interview concluded with open-ended re- flections on participantsâ ideal form of AI assistance and the barriers they perceived to adoption. This structured procedure allowed us to capture both in-the- moment behaviors during interaction and post-hoc reflections on onboarding, help-seeking, and workflow integration [56]. 3.5 Help Resources and Interaction Environment To preserve ecological validity, participants were explicitly encour- aged to use any resources they would normally draw upon in real- world tool exploration. Rather than restricting access to predefined materials, we sought to observe how participants engaged in self- directed help-seeking when confronted with uncertainty or system limitations [47]. As such, participants were free to consult the plat- formsâ built-in tutorials or tool-tips (as shown in Figure 1), external video tutorials and online documentation, community forums, or general-purpose search engines such as Google. We also allowed participants to make use of external AI tools (e.g., ChatGPT) during the study. Credit consumption was tracked through platform logs, and participant reactions to credit depletion or credit awareness were closely noted. These reactions often surfaced spontaneously in the course of interaction, particularly among casual users, and shaped the depth of their exploration, their reliance on external help resources, and their overall satisfaction with the generated models. This approach allowed us to capture credit-related concerns Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain as they naturally emerged, rather than foregrounding them as an explicit study variable. 3.6 Data Collection and Recording We employed a multimodal data collection strategy to capture both participantsâ direct interactions with the tools and their reflective accounts of the experience. Each session was recorded in full, includ- ing participantsâ screens, audio, and video feeds, with prior consent. The screen recordings documented fine-grained interaction details such as prompt edits, help seeking approaches, and inspection of generated models, while the audio channel preserved concurrent think-aloud commentary as well as the post-task interviews. In addition to recordings, the first author took notes during the sessions to document contextual observations, for example, moments of hesitation, nonverbal cues, or spontaneous remarks that might not be fully captured in transcripts. Where participants used ChatGPT to refine or troubleshoot their prompts, we archived those conversational exchanges to trace instances of cross-tool chaining. Platform-level logs were also collected when available, providing records of input prompts, generated outputs, and credit consumption. All audio recordings were transcribed and combined with the screen and system logs to provide a comprehensive dataset of user behavior. The observer notes and ChatGPT logs offered addi- tional contextual layers. Together, these complementary sources allowed us to triangulate across interactional, system-level, and self-reported data, strengthening credibility by comparing what participants said, what they did, and how the system responded. This triangulation reflects an interpretivist orientation, treating multiple perspectives as complementary in constructing meaning rather than as checks on a single truth [11]. 3.7 Data Analysis We analyzed the data using reflexive thematic analysis following Braun and Clarkeâs [11] six-phase method to identify patterns in how participants interacted with generative 3D modeling tools and how these patterns varied across expertise levels and credit conditions. Two of the authors collaboratively analyzed all transcripts, screen recordings, and field notes using a reflexive thematic analysis ap- proach. We iteratively refined the codebook through repeated dis- cussion, balancing descriptive accounts of user actions with inter- pretive insights into participantsâ reasoning, learning strategies, and expectations. To establish shared understanding of code bound- aries early in the process, both researchers independently coded an initial calibration subset of the data. For this subset, we calculated Cohenâs kappa (0.80) as a pragmatic check on alignment in code application, not to claim objectivity or coder neutrality. Disagree- ments were discussed reflexively and used to deepen interpretation and refine the analytic lens rather than treated as reliability failures. The remaining data were then coded through iterative, collaborative analysis consistent with reflexive thematic analysis practices. Codes were then clustered into broader themes such as AI chain- ing (e.g., using ChatGPT to refine prompts), trial-and-error adapta- tion and its constraints, perceptions of model quality and production readiness, expertise-driven workflow differences, trust breakdowns, and help-seeking avoidance in favor of self-guided exploration. Light quantification, such as counting participants who used ChatGPT or avoided help resources, was used to contextualize the prevalence of these behaviors within our qualitative themes. 3.8 Ethical Considerations and Compensation The study was approved by the Institutional Research Ethics Board. All participants were informed of their right to withdraw at any point. Personally identifiable data was removed from all recordings and transcripts. Participants received a $15 gift card as a token of appreciation for their time. 4 Results 4.1 Initial Mental Models: Prompting vs. Seeking Guidance All participants (14C, 12P) began the first task by opening their assigned generative 3D modeling tool. Their strategies diverged im- mediately: some entered a self-formulated description directly into the prompt field, while others consulted external resources, such as ChatGPT, Google Search, or the toolâs documentation, before pro- viding their first input. These early choices reflected participantsâ assumptions about how the system should be used to generate output. 4.1.1Beginning with Self-Formulated Prompts as the Default Strat- egy. Even though none of our participants had used generative 3D tools before, most participants (18/26) began their first task by entering their own prompt directly into the tool. This strategy was observed among half of casual users (7C/14C) and almost all pro- fessionals (11P/12P). Rather than consulting an external resource or built-in tips, participants moved directly to the blank text box to compose a descriptive input prompt. Their behavior suggests that the text box was interpreted as an immediate call to action, and that these participants felt confident starting without additional guidance. This behavior contrasts with prior studies (e.g., [47]) that showed that casual users often first turned to Google search and YouTube videos when learning an unfamiliar 3D modeling tool. In our study, however, none of the casual participants sought out YouTube or other video-based tutorials before generating their first prompt. One participant explicitly articulated this shift in mindset compared to how he would learn other types of software, stating, âWell, AI tool is made for everyone, and especially for non-professional use, so... they usually tend to be beginner-friendly, so I donât need any tutorialsâ (C15). A small subset (4C) consulted the official documentation, mainly to familiarize themselves with the interface. However, they quickly abandoned it (within a few seconds) when it failed to provide ac- tionable guidance. Among participants who began by entering prompts directly into the interface, casual and professional users showed stark differences in specificity. Professionals drew on prior CAD and 3D modeling experience to produce detailed, constraint-oriented prompts, often including dimensions, structural references, or geometry specifi- cations (see Table 2). Casual participants, by contrast, relied on shorter, less precise descriptions, or phrased conversationally. An CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana Table 2: Variations in prompting styles between professional and casual participants. Professional prompts were generally more structured, precise, and detailed, often specifying dimensions, materials, and design constraints, whereas casual prompts tended to be simpler, less specific, and conversational in tone. The average word length for a professional prompt was 83 vs for casuals it was 35. Professional PromptsCasual Prompts P21: Generate a glass coffee table with the legs made out of wood. The legs are Xs where the bottom of the X are the four feet of the table, and the top supports the glass of the table. The glass of the table is a rectangle with the corners rounded enough to be a full circle. The cross brace is a single piece with roughly 1x4x20 dimensions and is in the center of the X of the leg braces. C04: a table with elliptical glass top and wooden legs in the shape of an X P23: suppose you are building a table with glass top with it having rounded corners with some Radius have crossed wooden legs at both sides, make it simple with glossy top and ensure that these legs are connecting with each other C05: can you help me with an image of coffee table- which consists of wooden legs, and a glass top. specifically, there should be 4 wooden legs and a semi circled glass top. P16: I want to generate a table in which the table top material is glass, both semi-transparent and the legs are two parallel cross shapes with an element connecting them. these are all light coloured wood with square section C10: image of a coffee Table with wooden frame and glass top analysis of word counts of prompts across participants showed that casualsâ prompts averaged 35 words, while professionalsâ averaged 83 words (almost 2.4 times more). This gap reflected not only greater structural detail but also professionalsâ use of domain-specific vo- cabulary and explicit design constraints. 4.1.2 Seeking Guidance for Prompt Formulation. The remaining eight participants (7C, 1P) adopted a help-first approach, seeking guidance for input formulation before attempting to prompt the tool directly. Their strategies centered around three resources: external LLMs (ChatGPT), the toolâs official documentation on the web, and in-application help features. Use of ChatGPT: A subset of participants (3C, 1P) engaged in what we term AI chaining - the use of an LLM such as ChatGPT to generate or refine prompts for another prompt-based generative tool. Some saw ChatGPT as a faster, more articulate way to produce inputs for generative 3D tasks, âChatGPT can describe it faster and better than I can... in a shorter amount of timeâ (C03), while others used only after struggling with self-formulated prompts: âI try [to] create the prompt on my own. And if it doesnât work... I go to ChatGPTâ (C12). Another participant acknowledged this difficulty directly: âI will try [to prompt] but itâs a little challeng[ing] to describe what those structure[s are]â (C13). Official Documentation: Four participants (4C) began by con- sulting the tool documentation to learn input format, feature avail- ability, or relevant vocabulary. For some, this was a preparatory step before prompting, but most found the materials unhelpful or inadequate. As C08 remarked: ânot descriptive enough... I have something in my head, but I canât find it over hereâ. Similarly, C02 abandoned the documentation after failing to find usable guidance. Two participants (2C) shifted quickly to self-prompting, while two casuals (C02, C08) engaged in multi-help sequences, combining documentation and ChatGPT. In both cases, documentation shaped expectations, but ChatGPT supplied the actual text submitted to the tool. Built-in Help: Notably, no participants used the in-application help features. For example, in Meshy, none of the participants no- ticed or engaged with the hints icon in the UI (see Figure 1). Instead, participants either relied on ChatGPT or their own self-directed ex- perimentation. This pattern aligns with the active user paradox [14], in which software users, eager to begin their task, bypass available guidance even when it could improve performance. We revisit this point in the Discussion to consider its implications for the design of generative tools. 4.2 Differences in Trial and Error Strategies for Iteration After their first attempt, participants moved into cycles of trial and error, but their strategies diverged sharply by expertise. Casual users approached iteration as open-ended exploration, tweaking small descriptive details, layering on adjectives, or switching be- tween resources such as Google search and documentation when re- sults were unsatisfactory. Professionals, in contrast, pursued fewer targeted refinements grounded in prior CAD and 3D modeling knowledge. They refined prompts around specific parameters such as dimensions, geometry, or material constraints, emphasizing effi- ciency and structural precision. 4.2.1Direct In-Application Iteration and refinements vs AI Chaining. As described above, AI chaining was commonly used by a subset of participants before submitting prompts in the 3D modeling tool. In Task 1, 4 participants (3C, 1P) described the approach as a âfasterâ way to âget the phrasing rightâ without spending excessive credits on trial generations. Only 1 casual participant used ChatGPT in Task 2, and three participants(3C, 1P) employed it in Task 3. Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain Figure 3: Workflow visualizations of three participants illustrating the distinct approaches to onboarding and iterating with the generative 3D modeling tool. Left (P06âProfessional): Linear path where the the user did not take any external or in-built help. Center (P10âCasual): The casual user cycled across multiple help-seeking resources, including tool documentation (docs), Google search engines (GSE), while iteratively generating prompts and evaluating outputs. This pathway demonstrates a âtrial-and-errorâ orientation, with heavy reliance on external aids to understand of both prompting and tool capabilities. Right (P08âCasual): An AI-chaining approach where the participant depended on ChatGPT to draft and refine prompts before using them in the generative application, highlighting how casual users offloaded the burden of prompt crafting to external AI. Together, these three trajectories capture the major approaches in our study: (1) Self Formulated prompts, (2) Exploration scaffolded by external documentation and search, and (3) AI-for-AI prompting, where users relied on generative models to learn how to interact with other generative systems. In contrast, professionals largely avoided external AI support, expressing either distrust in its effectiveness, calling it "unneces- sary," or frustration with the quality âChatGPT sucks for 3D designâ (P21, professional). This reflected a broader sentiment that language- model-driven assistance lacked the precision and domain specificity required for complex 3D tasks: âItâs easier for me to put and describe things with words because I had something in mindâ (P16). This preference for self-reliance mirrors established patterns in expert creative practice, where efficiency, control, and trust in oneâs own skills outweigh perceived benefits of automated support. When participants did iterate, they pursued two main strategies: (1) prompt modification - making incremental changes to the original text; and (2) prompt restarting - abandoning the first attempt and writing a completely new description. In Task 1, many participants (7C, 7P) wrote the second prompt without relying on external resources, often adding specific descrip- tors or reordering attributes to prioritize certain features. Casual users typically added visual adjectives in the hope of nudging the AI toward aesthetic fidelity while professionals specified structural constraints table 2. The choice between modification and restart reflected how fail- ures were diagnosed. Casuals typically perceived flaws as âstylisticâ and were more likely to edit; professionals identified structural inaccuracies or mesh errors discarded the iteration itself. This dif- ference in iteration pathways reflects contrasting expectations: for casuals, it was a creative steering exercise, whereas for profession- als, it was a corrective precision task, and in many cases, not worth pursuing if the underlying model quality was judged insufficient: âall of these AI generation tools...generate everything from...scratch. They are using our first prompt...in some cases, we just want to have some changes in some part of the model.(P18)â Here, iteration was framed less as exploration and more as correction. 4.2.2Credit-Bound Behaviour and Iteration Risk. Credit-based gen- eration limits shaped the trial-and-error strategies of multiple par- ticipants, particularly casual users. Although the modeling tools provided a finite number of free generations, several participants treated credits as scarce resources, conserving them for âhigh- confidenceâ prompts rather than exploratory adjustments. Some stopped after a single attempt, even when results were imperfect: âIâm just trying to see like what type of prompts people use for text CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana ApproachInitial Onboarding ActionHelp-Seeking Strategies Casual (%)Professional (%)Casual (%)Professional (%) No Help Taken50.0%91.7%42.9%64.6% ChatGPT21.4%8.3%8.9%2.1% Tool Documentation28.6%0%8.9%2.1% Google Search0%0%19.6%6.2% Tutorial Videos0%0%3.6%2.1% Table 3: Comparison of help-seeking approaches across all tasks during the initial onboarding actions and subsequent trial- and-error iterations. Casual users often relied on external AI support (e.g., ChatGPT) and some consulted documentation; professionals mostly began with self-formulated prompts and iterated independently. Unlike prior HCI findings (e.g., [47]) where video tutorials were a dominant onboarding resource, our participants rarely turned to them. Instead, prompt-based onboarding itself emerged as a central help-seeking strategy, with users treating iterative prompt refinement as both a way to explore system capabilities and as a substitute for formal tutorials. to image prompts...I donât wanna waste my creditsâ (C11); âIf I click it, will it eat up more credits?â (C02). This often led to premature task closure, with participants accepting suboptimal outputs to preserve credits for later tasks. This behavioral shift was also re- flected quantitatively in participantsâ iteration patterns. In Task 1, participants who were not under credit constraints averaged 4.4 prompt iterations per task, compared to only 2.3 iterations among those who were credit-bound. This suggests that exploration de- creased noticeably once participants became aware of credit risk. This pattern intensified in later tasks: by Task 2, 12 participants (6C, 6P) made only a single attempt, and in Task 3, 13 participants (9C, 4P) demonstrated the same one-attempt behavior. Professionals were also credit-aware, but their reluctance stem- med from doubts about the toolâs ceiling of quality rather than credit scarcity. As P16 explained: âI donât think Iâm gonna use them [credits], even if theyâre free, because my time also has value... if Iâm not getting anything even close to what Iâm desiring, thereâs no point.â For both groups, iteration decisions reflected resource- rational tradeoffs, where expectations about tool limits, rather than actual capacity, shaped willingness to invest effort. Rather than empirically testing the systemâs potential, partici- pants acted on their mental models of what the tool could or could not deliver: casual users assumed that additional credits were un- likely to be âworth it,â while professionals assumed that the AIâs output ceiling was inherently too low. This aligns with broader cognitive theories of resource-rational decision making and satis- faction, where users conserve effort (or credits) once their expecta- tions about system limits are set, even if those expectations may underestimate the toolâs actual capabilities [7, 19]. 4.2.3 Help-Seeking Patterns. External help-seeking was surpris- ingly sparse given that participants were encouraged to use multiple support channels, including built-in documentation, online tutori- als, and general web resources (see Table 3). Google searches were used by six casuals and three professionals across all tasks, mostly to identify terminology (e.g., âvertices,â âcross brace dimensionsâ). YouTube was even rarer (two casuals total), with participants noting that videos were too slow for time-limited tasks. Official documen- tation was also underused, accessed by only four participants in Task 1 and fewer thereafter, often via Google rather than the toolâs own interface. Participants made comments such as, âif I need some- thing urgently, I donât have time to go and learn the resourcesâ (C08). This number dropped to 3 casuals in Task 2 and 2 participants (1 casual, 1 professional) in Task 3. A common complaint was the unfamiliarity with the jargon: âI have something in my head, but I canât find it over here [documentation] I donât know what is what.â (P08) Another participant expressed the hesitation to use tutorials: âI feel like that the tutorial section was more about... how to make certain things, but I kind of wanted...more general bit, like how do I use the different aspects of the AI toolâ (C20). Casual users were more inclined to experiment with layered help-seeking, often moving between multiple resources within the same task. For example, in Task 2, C02 sequentially consulted the Spline documentation, then searched Google, moved to YouTube, and finally used ChatGPT to refine their prompt after repeated mismatches with the target model. C12âs narrative captures the sentiment common among casuals: âYoutube tutorials are helpful when youâre a novice, and you just want a general walk through, an orientation and have someone explain everything to you. But also AI tools like Chatgpt are really helpful if youâre stuck on a specific scenario, or Iâm trying to do this exact thing in my project, and I donât know where to click this or that it can actually give you pretty good step-by-step instructionsâ In contrast, professionals showed minimal reliance on external resources (only 3 professionals used Google at any point). These users expressed confidence in their own ability to diagnose and address modeling issues, even when confronted with unsatisfac- tory AI outputs. Several described external resources as too slow or not specific enough for their precision-driven needs, preferring instead to either restart from scratch or revert to traditional CAD workflows. 4.3 Differences in Workflow Perceptions Participantsâ interpretations of AI-generated outputs revealed a clear divide in how casual and professional users: (1) evaluated AI outputs, and, (2) where they positioned the tool within their broader workflows. Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain (a) Generalized Workflow of the Casual Participants (b) Generalized Workflow of the Professional Participants Figure 4: Generalized workflows of participants. (a) Casual participants often treated generated models as immediately ready for 3D printing, overlooking issues such as surface anomalies or structural stability. (b) Professional participants followed two distinct pathways: a post-editing pathway, where they repaired or refined AI-generated models before considering them print-ready, and a model-from-scratch pathway, where they rejected the AI output, returned to ideation, and created new models for downstream use. These contrasting workflows illustrate how casual users emphasized immediacy and usability, while professionals foregrounded production quality and long-term integration. 4.3.1 Evaluation Standards: Good Enough vs. Production-Ready. Casual participants (n = 14) often treated generated outputs as finished artifacts, suitable for immediate 3D printing or display. For them, visual resemblance was the dominant quality metric: if the model âlooked likeâ the intended object, it was assumed to be fabrication-ready, regardless of mesh integrity, wall thickness, or structural stability. As C12âs remarked: âI was relieved at how easy it was to create an object just with AI prompts... usually you have to get really familiar with tools, learn how to create stuff from scratch.â Similarly, C08 found that getting an "approximate" match was sufficient: âIt was able to get the basic idea in check... even though it might not be 100% similar, it was good enough so I could convey my idea.â Professionals (n = 12), in contrast, consistently judged outputs as not production-ready in their unedited form. Professionalsâ as- sessments focused on structural quality and fabrication feasibility, systematically inspecting geometry for errors (e.g., non-manifold meshes, thin supports, or irregular topologies). P21 was blunt: âThis is not usable in a final product at all.â P17 highlighted issues of print- ability: âThese smaller details around here? It might make 3D printing hard.â P06 emphasized watertight geometry: âTo have anything 3D print ready... there should be no thinning or holes. Like, it should be a solid model. With this, I donât think it is.â 4.3.2Workflow Positioning: Direct Output vs. Early-Stage Ideation. The differing evaluations of the output of our participants shaped distinct workflow trajectories. Casuals often pursued a minimal sequence of promptâprint, skipping intermediate steps such as inspection or refinement. As C08 explained, speed outweighed precision: âI think for simpler things, like... a cupcake... this is good. Itâs available widely. I can do it very fast.â The ability to bypass labor-intensive modeling became the main source of satisfaction, redefining what âgood enoughâ meant in their workflows. Professionals instead positioned the AI tool at the ideation stage of their workflow, where the outputs could serve as inspiration. Three participants (P09, P17 and P18) attempted post-editing (De- signâ3D ModelâPost-Editingâ3D Print), exporting models into advanced 3D modeling tools, such as Blender, SolidWorks, or Rhino, for hole filling, edge merging, and wall-thickness adjust- ments. However, this was time-consuming and prone to new errors (e.g., corrupted UV maps). The more common pattern (9 out of 12 professionals) was to abandon editing entirely, following Designâ 3D ModelâIdeationâModel from Scratchâ3D Print. As P17 noted: âIf itâs like bumpy, and I have to clean it up myself, that would be kind of annoying.â P21 similarly emphasized that AI benefits only at the earliest stages: â only saves me time in the fastest ideation portions.â This distinction echoes prior work on casual makers in walk- up-and-use 3D printing environments [39], where casuals often avoided refinement altogether. In contrast, for professionals, AI out- puts offered conceptual value but rarely fabrication utility. In both cases, workflow breakdowns curtailed deeper design exploration, casuals due to usability hurdles, and professionals due to tool-chain complexity. 5 Discussion 5.1 Larger Context of Our Findings Our study reveals a notable shift in how users begin learning and seeking help with creative software: rather than relying on videos, forums, or search engines as documented in prior HCI work [34,47,58], participants increasingly turned to other generative AI CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana Table 4: Key Insights of casual and professional participants engaged with generative 3D modeling tools. The comparison highlights distinct patterns across initial engagement, prompt style, help-seeking behavior, iteration strategies, workflow integration, and output evaluation, reflecting the contrasting approaches and expectations shaped by varying levels of expertise. DimensionCasualsProfessionals Initial EngagementBegin with a quick, self-formulated prompt or seek help from ChatGPT or documentation before starting. Usually start with a prompt-first approach, relying on their own descriptive skills without consulting help. Prompt StylePrompts tend to be shorter, less structured, and more visually descriptive. Prompts are highly detailed, structurally explicit, and specify precise geometric or dimensional constraints. Help-SeekingFrequently use layered help sources like ChatGPT, Google, and documentation, often in ad-hoc sequences. Rarely seek external help, preferring self-reliance and domain expertise over outside resources. Prompt IterationMore likely to make small wording tweaks or stylistic edits, sometimes chaining AI tools for refinement. Prefer major rewrites or start from scratch when out- puts deviate structurally, avoiding external AI tools. Workflow IntegrationTreat AI-generated outputs as production-ready, mov- ing directly from generation to intended use. Use outputs as early prototypes or concept references, requiring significant post-editing or full remodeling. Output EvaluationJudge quality mainly by visual similarity to the intended object, often overlooking structural flaws. Assess outputs against strict manufacturability stan- dards, identifying geometric and structural issues be- fore use. systems to understand or operate the primary 3D tool. Casual partic- ipants, in particular, rarely explored Meshy or Spline through their interfaces; instead, they moved directly into AI chaining workflows for instance, asking ChatGPT to rewrite prompts, fix prompt phras- ing, or explain unexpected model outputs before returning to the primary 3D tool. Professionals engaged in less external AI use but still relied on in-application AI features for prompting and making trial and error rather than documentation or tutorials. Taken to- gether, these patterns suggest a key shift in onboarding from feature exploration [21,48,49] toward prompt formulation as the primary entry point. As more modeling platforms integrate embedded copi- lots [6], model-assisted refinement [54], or auto-complete designs [18], we expect this GenAI-to-GenAI help ecosystem to become an increasingly dominant mode of onboarding and problem-solving. Despite these shifts, our findings also show continuity with long- standing theories of how users learn and interact with feature-rich software. Participants skipped tutorials and avoided in-built tool documentation in favor of immediate action, consistent with min- imalist instruction [13] and the paradox of the active user [14], which predicts that users prioritize âgetting startedâ over learning even when it leads to inefficiency. This preference for immediate action was reinforced by production bias, the tendency to favor quick progress over deeper understanding and structured learn- ing [14], which was especially visible among casual users who aimed to obtain a usable model as quickly as possible, even if that meant relying on imperfect prompts or overlooking system features. Likewise, usersâ help-seeking patterns reflect well-documented self- regulated learning processes [2,32]: users monitored breakdowns, evaluated failed outputs, and selectively turned to external assis- tance, often alternating between instrumental help-seeking (seeking understanding, more common in professionals) and executive help- seeking (seeking fast solutions, common among casuals). Taken together, our findings suggest that while generative AI introduces new mechanisms such as AI chaining and language-based onboard- ing, the underlying motivations that shape usersâ software learning and help-seeking behavior remain well-accounted for by earlier conceptual and theoretical observations in HCI. 5.2 Help-Seeking as Multi-AI Ecosystem Behavior Our findings suggest a reconfiguration of help-seeking for feature- rich software, one defined not by consulting manuals or tutorials but by coordinating assistance across multiple AI systems [44]. Rather than relying on a single source of support, participants treated different generative tools as complementary resources, each offering partial guidance or corrective feedback. AI Chaining was not a workaround but an intentional strategy shaped by credit limits, uncertainty in outputs, and the opacity of individual models. These patterns suggest that HCI should increasingly understand help-seeking as a distributed process and treat AI tools not as standalone systems but as interdependent collaborators within user-constructed help ecosystems. 5.2.1 Design Implication: Designing Workflow-Level Multi-Model Guidance. Our findings suggest that Gen AI tools should no longer treat the creation pipeline as a single opaque step, but instead sur- face the underlying workflow in a way that supports both learning and expert control. By integrating this behavior into the interface, tools could make each stage of the generative pipeline visible and optionally editable, allowing novices to follow a guided, progres- sive path (generateârefineâtextureâexport) while enabling experts to collapse, bypass, or reorder stages. This approach could build upon established principles of progressive disclosure and staged workflows in complex software [28,34] , as well as dis- tributed cognition perspectives showing that users naturally of- fload cognitive work across multiple artifacts when solving complex Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain tasks [36]. Rather than hiding intermediate steps, systems could offer concise explanations of what each model is doing, show in- termediate artifacts, and provide diagnostic feedback at transition points helping users build accurate mental models of the process. Such workflow-level guidance also creates a foundation for explicit handoffs between AI tools, acknowledging the reality of multi- model use and reducing the trial-and-error burden users currently face. By making the generative pipeline transparent and adaptable, tools can better support the very different expectations casual and professional users bring to AI-assisted 3D modeling. 5.3 Rethinking Software Onboarding in the Age of Prompts Instead of engaging with tutorials, interface walkthroughs, or fea- ture exploration, we found that users begin by producing language instructions. This pattern shows a shift in onboarding from learning how to operate a tool to learning how to express intentions in a way the AI can interpret. This reframes onboarding as a negotiation with the languageâaction gap of generative models rather than the interface itself [46,69]. For casual users, this shift lowers the barrier to âgetting started,â because the entry point becomes conversational rather than procedural. For professionals, however, this inversion complicates expectations of precision, reproducibility, and control qualities traditionally grounded in direct manipulation workflows. Professionalsâ expertise, built around spatial reasoning, constraint management, and iterative refinement, does not always transfer into a text-based interaction model. Together, these patterns il- lustrate a new form of software learning: one where the primary challenge is not feature discovery but understanding how to collab- orate with a stochastic system whose internal logic is opaque. This creates a divergence in experience: beginners benefit from reduced complexity, while experts encounter friction when the model fails to match their expectations for accuracy, parameter control, or workflow alignment. 5.3.1 Design Implication: Designing Micro-Scaffolds That Lever- age the Prompt Loop. Traditional contextual help has existed since early intelligent tutoring systems [2], but generative AI creates an opportunity to design in-time micro-scaffolds that work within the userâs existing prompt-iteration workflow rather than interrupt- ing it. Instead of surfacing generic hints or feature explanations, systems could provide prompt-level scaffolding that adapts to the userâs pattern of failures. For instance, they could surface example prompts when the model detects repeated semantic misunderstand- ings, provide parameter suggestions when outputs show consistent geometric issues, or highlight generative constraints (âthin sur- facesâ, âunsupported geometryâ) as lightweight annotations on the prompt itself. This idea could further build on research showing that productive help is most effective when it is embedded in the userâs immediate task state [2] and when it supports usersâ existing strategies rather than requiring entirely new ones. By designing scaffolds that amplify the userâs natural âlearn-by-promptingâ loop rather than attempting to reintroduce traditional tutorial models, tools can better support both casual users who rely on rapid ex- perimentation and professionals who require conceptual clarity to maintain precision. In this sense, the implication is not simply to âadd contextual help,â but to rethink help as prompt-guidance infrastructure, a direction aligned with adaptive interfaces [28] and modern HCI research that treats AI as both collaborator and instructor [41]. 5.4 âGood Enoughâ vs âNot Even Closeâ: Divergent Standards of Sufficiency of AI Output Our findings reveal a clear divergence in how casual and profes- sional users evaluate the sufficiency of generative 3D outputs. Ca- sual users frequently accepted visually plausible models as âgood enough,â collapsing the traditional design pipeline into a direct path from concept to artifact. In contrast, professionals judged the same results as structurally inadequate or unsuitable for downstream workflows, identifying issues such as non-manifold geometry, miss- ing supports, or inaccuracies that compromise printability. This division echoes prior observations in computer graphics and vision that generative models optimize for visual plausibility rather than geometric correctness [75], and that mesh-quality issues remain a persistent challenge even in state-of-the-art 3D generation pipelines [86][42]. By examining user workflows and help-seeking patterns, our study shows that âvisual realismâ valued in generative AI re- search does not translate uniformly to âworkflow-ready qualityâ valued by professionals. This underscores why human-centered evaluation is essential: while generative 3D lowers barriers for new- comers, it also widens the experiential gap between novices and experts by privileging immediacy over precision. Designing sys- tems that support both groups within the same interface remains an open challenge, requiring tools that recognize divergent stan- dards of sufficiency rather than treating output quality as a single, universal metric. 5.4.1 Design Implication: Designing Dual Modes for Concept vs Production Workflows. Our findings suggest that generative 3D modeling tools would benefit from explicitly supporting two paral- lel modes: one optimized for rapid, visually plausible concept gen- eration and another designed for structurally reliable, production- ready modeling. Introducing a Concept Mode would embrace these strengths by optimizing for immediacy, creative exploration, and low cognitive overhead, echoing prior HCI work showing that novices benefit from simplified workflows, reduced parameter ex- posure, and rapid feedback [28,34]. In contrast, a Production Mode could surface the kinds of controls and diagnostics that experts expect in precision-oriented environments, such as mesh manifold checks, thickness warnings, printability indicators, and parametric constraints, aligning with practices from CAD, animation, and fab- rication pipelines where structural integrity and export fidelity are critical [4]. Supporting these two modes allows systems to respect the divergent standards of sufficiency of AI output among casual and professional users. Prior HCI studies on user interaction with feature-rich software show that when interfaces cater exclusively to novices, experts experience friction and inefficiency [26,47]. In contrast, enabling users to fluidly switch between conceptual and production workflows acknowledges both ends of the expertise spectrum without forcing a one-size-fits-all interaction model. Such dual-track design could also offer a more meaningful interpretation CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana of âdemocratization,â ensuring that generative tools empower new- comers without compromising the rigor necessary for professional practice. 5.5 Credit Constraints and the Dynamics of Generative Exploration We found that credit-based constraints significantly shaped how users explored generative 3D tools, especially among casual users who became risk-averse and limited iteration to avoid âwastingâ credits. Although our study focused on 3D modeling, this pattern could generalize to other generative domains where outputs are token-metered or computationally expensive (e.g., AI-assisted CAD, texture generation, and parametric design). Our study suggests that credit constraints amplify production bias, heighten loss aversion, and shape trust decisions, pushing users to judge whether an out- put is âworth a creditâ based on surface plausibility rather than structural correctness. Nevertheless, domain-specific constraints may influence how strongly these behaviors appear. 5.5.1 Design Implication: Designing for Credit Transparency and Expectation Calibration . To counter credit-driven constraints, gen- erative tools could make credit consumption more transparent by previewing the expected credit cost of an action, the likely fi- delity or structural reliability of the output, and lower-cost alterna- tives before generation occurs. Such predictive scaffolding echoes principles from trust-calibration literature, which emphasize that users form more resilient trust when systems communicate un- certainty, limitations, and probable failure modes upfront rather than retroactively [63]. For example, when a user attempts to gen- erate a mechanically precise or physically constrained object, the system might warn that generative models typically struggle with thickness tolerances, manifolds, or print-ready geometry, issues well-documented in 3D generative modeling research [42,86]. By framing credit cost and output reliability together, systems can help users make informed trade-offs between exploration and resource consumption. This is particularly important because, as our findings showed, uncertainty about credit consequences caused participants to prematurely settle for suboptimal outputs or rely on external AIs (e.g., ChatGPT) as a no-cost workaround. Transparent cost-benefit signals, therefore, not only reduce anxiety around credit spending but also support more deliberate exploration, encourage productive iteration, and calibrate user expectations toward what generative 3D models can and cannot reliably produce. Designing for such cal- ibrated expectations is essential for building trust without fostering overconfidence, a challenge repeatedly highlighted in recent work [43, 50] on AI-assisted creativity and decision-making. 5.6 Limitations Our study has several limitations that frame the scope of our find- ings. First, while the participant pool included both casual and professional 3D designers, our professional sample covered a range of roles such as product design, CAD engineering, and 3D artistry, but not all subdomains such as game asset modeling or large-scale fabrication. Second, the study focused on two specific generative 3D modeling tools. Although these platforms are representative of current systems, differences in algorithmic back-ends, mesh quality, and interface design could yield different user behaviors. Third, our task-focused study design may have influenced iteration behav- iors. In real-world practice, users often work under time, budget, and collaborative constraints, which could amplify or dampen the strategies observed. For example, professionalsâ may view post- editing in a commercial context where AI outputs could serve as rapid prototyping starting points under strict deadlines. Finally, our credit-based conditions reflected the default limits of the tools rather than an experimental manipulation; future work could vary credit availability more systematically to assess its impact on explo- ration and help-seeking. Additionally, the fixed task order may have introduced order effects. We prioritized a consistent progression across tasks to ensure comparability; however, this may have unin- tentionally amplified prompt-first behaviors early in the session and suppressed later exploration or help-seeking due to time pressure or credit depletion. As a result, some participants may have relied more heavily on initial prompting strategies rather than revisiting tutorials or alternative workflows later. Future work could counter- balance task order or vary credit availability to better isolate these effects. We made a proactive effort to recruit participants across gen- ders and age groups, though most ultimately reflected the younger demographic that currently dominates early adoption of generative 3D modeling tools. Some groups remain underrepresented due to the emerging nature of these systems and the characteristics of their present-day user base. As generative 3D tools mature and attract broader audiences, future work should expand demographic diversity to capture a wider range of perspectives. 6 Conclusions Our study shows how generative 3D modeling reshapes design by collapsing the boundaries between learning, creating, and seeking help. These shifts are less about skill than about different visions of AI as a shortcut, collaborator, or unreliable partner. Participants often orchestrated multiple AI systems in tandem, pointing to a future where help-seeking extends beyond manuals or forums into distributed dialogues across AI tools. For designers of creative tech- nologies, this means building systems that not only generate results but also adapt to diverse mindsets, supporting both users seeking ready-made outputs and those treating AI as a design partner. For HCI, the broader challenge is to see AI ecosystems as emerging sites of creative practice where expertise, authorship, and support are being redefined. Attending to these transformations is vital if HCI is to keep pace with how generative AI reshapes everyday design work. Ultimately, design is no longer solitary; it happens in concert with AI. 7 Acknowledgments We thank the Natural Sciences and Engineering Research Council of Canada (NSERC). References [1] Adobe, Inc. 2025. Adobe Firefly. The ultimate creative AI solution. https://w. adobe.com/products/firefly.html. [2] Vincent Aleven, Elmar Stahl, Silke Schworm, Frank Fischer, and Raven Wallace. 2003. Help Seeking and Help Design in Interactive Learning Environments. Re- view of Educational Research 73, 3 (2003), 277â320. doi:10.3102/00346543073003277 [3]Narges Ashtari, Parsa Alamzadeh, Gayatri Ganapathy, and Parmit Chilana. 2022. PONI: A Personalized Onboarding Interface for Getting Inspiration and Learn- ing About AR/VR Creation. In Nordic Human-Computer Interaction Conference Rethinking Onboarding and Help-Seeking with Generative 3D Modeling ToolsCHI â26, April 13â17, 2026, Barcelona, Spain (Aarhus, Denmark) (NordiCHI â22). Association for Computing Machinery, New York, NY, USA, Article 32, 14 pages. doi:10.1145/3546155.3546642 [4]Marco Attene. 2010. A lightweight approach to repairing digitized polygon meshes. The Visual Computer 26 (11 2010), 1393â1406. doi:10.1007/s00371-010- 0416-3 [5]Song Bai and Jie Li. 2024. Progress and Prospects in 3D Generative AI: A Technical Overview including 3D human. arXiv:2401.02620 [cs.AI] https://arxiv.org/abs/ 2401.02620 [6] Shraddha Barke, Michael B. James, and Nadia Polikarpova. 2022. Grounded Copilot: How Programmers Interact with Code-Generating Models. arXiv:2206.15000 [cs.HC] https://arxiv.org/abs/2206.15000 [7]Rahul Bhui, Lucy Lai, and Samuel J Gershman. 2021. Resource-rational decision making. Current Opinion in Behavioral Sciences 41 (2021), 15â21. doi:10.1016/j. cobeha.2021.02.015 Value based decision-making. [8]Blender Foundation. 2025. Blender â The Free and Open Source 3D Creation Software. https://w.blender.org/. [9]Joel Brandt, Mira Dontcheva, Marcos Weskamp, and Scott R. Klemmer. 2010. Example-centric programming: integrating web search into the development environment. In Proceedings of the SIGCHI Conference on Human Factors in Com- puting Systems (Atlanta, Georgia, USA) (CHI â10). Association for Computing Machinery, New York, NY, USA, 513â522. doi:10.1145/1753326.1753402 [10] Joel Brandt, Philip J. Guo, Joel Lewenstein, Mira Dontcheva, and Scott R. Klemmer. 2009. Two studies of opportunistic programming: interleaving web foraging, learning, and writing code. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Boston, MA, USA) (CHI â09). Association for Computing Machinery, New York, NY, USA, 1589â1598. doi:10.1145/1518701. 1518944 [11] Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psy- chology. Qualitative Research in Psychology 3, 2 (2006), 77â101. doi:10.1191/ 1478088706qp063oa [12]Leah Buechley and Benjamin Mako Hill. 2010. LilyPad in the wild: how hardwareâs long tail is supporting new engineering and design communities. In Proceedings of the 8th ACM Conference on Designing Interactive Systems (Aarhus, Denmark) (DIS â10). Association for Computing Machinery, New York, NY, USA, 199â207. doi:10.1145/1858171.1858206 [13]John M. Carroll (Ed.). 1998. Minimalism Beyond the Nurnberg Funnel. The MIT Press, Cambridge, MA. doi:10.7551/mitpress/4616.001.0001 Published in association with the Society for Technical Communication. [14] John M. Carroll and Mary Beth Rosson. 1987. Paradox of the active user. MIT Press, Cambridge, MA, USA, 80â111. [15] Antonio A. Casilli. 2024. Digital Labor and the Inconspicuous Production of Artificial Intelligence. arXiv:2410.05910 [cs.CY] https://arxiv.org/abs/2410.05910 [16]Edwige Chauvergne, Martin Hachet, and Arnaud Prouzeau. 2023. User Onboard- ing in Virtual Reality: An Investigation of Current Practices. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI â23). Association for Computing Machinery, New York, NY, USA, Article 711, 15 pages. doi:10.1145/3544548.3581211 [17]Tianrun Chen, Chenglong Fu, Lanyun Zhu, Papa Mao, Jia Zhang, Ying Zang, and Lingyun Sun. 2023. Deep3DSketch: 3D modeling from Free-hand Sketches with View- and Structural-Aware Adversarial Training. arXiv:2312.04435 [cs.M] https://arxiv.org/abs/2312.04435 [18]Xiang âAnthonyâ Chen, Jeeeun Kim, Jennifer Mankoff, Tovi Grossman, Stelian Coros, and Scott E. Hudson. 2016. Reprise: A Design Tool for Specifying, Gen- erating, and Customizing 3D Printable Adaptations on Everyday Objects. In Proceedings of the 29th Annual Symposium on User Interface Software and Technol- ogy (Tokyo, Japan) (UIST â16). Association for Computing Machinery, New York, NY, USA, 29â39. doi:10.1145/2984511.2984512 [19]Thomas J. Chermack. 2003. Mental Models in Decision Making and Implications for Human Resource Development. Advances in Developing Human Resources 5, 4 (2003), 408â422. doi:10.1177/1523422303257373 [20]Parmit K. Chilana, Amy J. Ko, and Jacob O. Wobbrock. 2012. LemonAid: selection- based crowdsourced contextual help for web applications. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Austin, Texas, USA) (CHI â12). Association for Computing Machinery, New York, NY, USA, 1549â1558. doi:10.1145/2207676.2208620 [21]Andy Cockburn, Carl Gutwin, and Saul Greenberg. 2007. A predictive model of menu performance. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (San Jose, California, USA) (CHI â07). Association for Com- puting Machinery, New York, NY, USA, 627â636. doi:10.1145/1240624.1240723 [22] Giovanni Emanuele Corazza, Sergio Agnoli, Ana Jorge Artigau, Ronald A. Beghetto, Nathalie Bonnardel, Irene Coletto, Angela Faiella, Katusha Gerardini, Kenneth Gilhooly, Vlad P. GlÄveanu, Michael Hanchett Hanson, Hansika Kapoor, James C. Kaufman, Yoed N. Kenett, Anatoliy V. Kharkhurin, Simone Luchini, Margaret Mangion, Mario Mirabile, Felix-Kingsley Obialo, Connie Phelps, Roni Reiter-Palmon, Jeb S. Puryear, Eleonora Diletta Sarcinella, Min Tang, Giulia Maria Vavassori, Florent Vinchon, Indre Viskontas, Selina Weiss, Dimitrios Zbainos, and Todd Lubart. 2025. Cyber-Creativity: A Decalogue of Research Challenges. Journal of Intelligence 13, 8 (2025). doi:10.3390/jintelligence13080103 [23]Cheryl Coyle and Mary Peterson. 2016. Learnability Testing of a Complex Software Application, Vol. 9747. 560â568. doi:10.1007/978-3-319-40355-7_53 [24]Jonathan Edwards. 2004. Example centric programming. SIGPLAN Not. 39, 12 (Dec. 2004), 84â91. doi:10.1145/1052883.1052894 [25]G.J Elliott, E Jones, and P Barker. 2002. A grounded theory approach to modelling learnability of hypermedia authoring tools. Interacting with Computers 14, 5 (2002), 547â574. doi:10.1016/S0953-5438(02)00021-8 [26]Barrett Ens, Tovi Grossman, Fraser Anderson, Justin Matejka, and George Fitz- maurice. 2015. Candid Interaction: Revealing Hidden Mobile and Wearable Com- puting Activities. In Proceedings of the 28th Annual ACM Symposium on User Inter- face Software & Technology (Charlotte, NC, USA) (UIST â15). Association for Com- puting Machinery, New York, NY, USA, 467â476. doi:10.1145/2807442.2807449 [27] Ziv Epstein, Aaron Hertzmann, the Investigators of Human Creativity, Memo Akten, Hany Farid, Jessica Fjeld, Morgan R. Frank, Matthew Groh, Laura Herman, Neil Leach, Robert Mahari, Alex âSandyâ Pentland, Olga Russakovsky, Hope Schroeder, and Amy Smith. 2023. Art and the science of generative AI. Science 380, 6650 (June 2023), 1110â1111. doi:10.1126/science.adh4451 [28] Leah Findlater and Joanna McGrenere. 2004. A comparison of static, adaptive, and adaptable menus. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Vienna, Austria) (CHI â04). Association for Computing Machinery, New York, NY, USA, 89â96. doi:10.1145/985692.985704 [29] Sarah E. Fox, Samantha Shorey, Esther Y. Kang, Dominique Montiel Valle, and Estefania Rodriguez. 2023. Patchwork: The Hidden, Human Labor of AI Integra- tion within Essential Work. Proc. ACM Hum.-Comput. Interact. 7, CSCW1, Article 81 (April 2023), 20 pages. doi:10.1145/3579514 [30] G. W. Furnas, T. K. Landauer, L. M. Gomez, and S. T. Dumais. 1987. The vocabulary problem in human-system communication. Commun. ACM 30, 11 (Nov. 1987), 964â971. doi:10.1145/32206.32212 [31]Narayana Gaddam. 2022. Generative AI for Creating Hyper-Realistic 3D Human Models. INTERNATIONAL JOURNAL OF INFORMATION TECHNOLOGY AND MANAGEMENT INFORMATION SYSTEMS 13 (06 2022), 137â149. doi:10.34218/ IJITMIS_13_01_013 [32]Sharon Nelson-Le Gall. 1985. Chapter 2: Help-Seeking Behavior in Learning. Re- view of Research in Education 12, 1 (1985), 55â90. doi:10.3102/0091732X012001055 [33]Google DeepMind. 2024. Google DeepMind and Gemini Generative AI. https: //deepmind.google. [34] Tovi Grossman, George Fitzmaurice, and Ramtin Attar. 2009. A survey of software learnability: metrics, methodologies and guidelines. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (Boston, MA, USA) (CHI â09). Association for Computing Machinery, New York, NY, USA, 649â658. doi:10. 1145/1518701.1518803 [35]Jeffrey Heer. 2019. Agency plus automation: Designing artificial intelligence into interactive systems. Proceedings of the National Academy of Sciences 116, 6 (Feb. 2019), 1844â1850. doi:10.1073/pnas.1807184115 [36]James Hollan, Edwin Hutchins, and David Kirsh. 2000. Distributed cognition: toward a new foundation for human-computer interaction research. ACM Trans. Comput.-Hum. Interact. 7, 2 (June 2000), 174â196. doi:10.1145/353485.353487 [37]Irene Hou, Owen Man, Kate Hamilton, Srishty Muthusekaran, Jeffin Johnykutty, Leili Zadeh, and Stephen MacNeil. 2025. "All Roads Lead to ChatGPT": How Generative AI is Eroding Social Interactions and Student Learning Communities. arXiv:2504.09779 [cs.CY] https://arxiv.org/abs/2504.09779 [38]Irene Hou, Sophia Metille, Zhuo Li, Owen Man, Cynthia Zastudil, and Stephen MacNeil. 2024. The Effects of Generative AI on Computing Studentsâ Help- Seeking Preferences. arXiv:2401.02262 [cs.HC] https://arxiv.org/abs/2401.02262 [39]Nathaniel Hudson, Celena Alcock, and Parmit K. Chilana. 2016. Understand- ing Newcomers to 3D Printing: Motivations, Workflows, and Barriers of Casual Makers. In Proceedings of the 2016 CHI Conference on Human Factors in Comput- ing Systems (San Jose, California, USA) (CHI â16). Association for Computing Machinery, New York, NY, USA, 384â396. doi:10.1145/2858036.2858266 [40]Amir Jahanlou and Parmit K Chilana. 2024. How Example-Based Authoring of Motion Graphics Impacts Creative Expression: Differences in Perceptions of Professional and Casual Motion Designers. In Proceedings of the 16th Conference on Creativity & Cognition (Chicago, IL, USA) (C&C â24). Association for Computing Machinery, New York, NY, USA, 347â357. doi:10.1145/3635636.3656197 [41]Tingting Jiang, Zhumo Sun, Shiting Fu, and Yan Lv. 2024. Human-AI interaction research agenda: A user-centered perspective. Data and Information Management 8, 4 (2024), 100078. doi:10.1016/j.dim.2024.100078 [42]Heewoo Jun and Alex Nichol. 2023. Shap-E: Generating Conditional 3D Implicit Functions. arXiv:2305.02463 [cs.CV] https://arxiv.org/abs/2305.02463 [43] Patricia K. Kahr, Gerrit Rooks, Martijn C. Willemsen, and Chris C. P. Snijders. 2024. Understanding Trust and Reliance Development in AI Advice: Assessing Model Accuracy, Model Explanations, and Experiences from Previous Interactions. ACM Trans. Interact. Intell. Syst. 14, 4, Article 29 (Dec. 2024), 30 pages. doi:10.1145/ 3686164 [44]Victor Kaptelinin and Bonnie Nardi. 2006. Acting with Technology: Activity Theory and Interaction Design. MIT Press, Cambridge, MA. [45]Nimisha Karnatak, Adrien Baranes, Rob Marchant, Huinan Zeng, TrĂona Butler, and Kristen Olson. 2025. Expanding the Generative AI Design Space through CHI â26, April 13â17, 2026, Barcelona, SpainKanak Gautam, Poorvi Bhatia, and Parmit K Chilana Structured Prompting and Multimodal Interfaces. arXiv:2504.14320 [cs.HC] https://arxiv.org/abs/2504.14320 [46]Anjali Khurana, Hariharan Subramonyam, and Parmit K Chilana. 2024. Why and When LLM-Based Assistants Can Go Wrong: Investigating the Effectiveness of Prompt-Based Interactions for Software Help-Seeking. In Proceedings of the 29th International Conference on Intelligent User Interfaces (IUI â24). ACM, 288â303. doi:10.1145/3640543.3645200 [47]Kimia Kiani, George Cui, Andrea Bunt, Joanna McGrenere, and Parmit K. Chilana. 2019. Beyond "One-Size-Fits-All": Understanding the Diversity in How Software Newcomers Discover and Make Use of Help Resources. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI â19). Association for Computing Machinery, New York, NY, USA, 1â14. doi:10.1145/3290605.3300570 [48]Amy J. Ko, Brad A. Myers, and Htet Htet Aung. 2004. Six Learning Barriers in End-User Programming Systems. In 2004 IEEE Symposium on Visual Languages - Human Centric Computing. 199â206. doi:10.1109/VLHCC.2004.47 [49]Ben Lafreniere, Tovi Grossman, Justin Matejka, and George Fitzmaurice. 2014. Investigating the feasibility of extracting tool demonstrations from in-situ video content. In Proceedings of the SIGCHI Conference on Human Factors in Comput- ing Systems (Toronto, Ontario, Canada) (CHI â14). Association for Computing Machinery, New York, NY, USA, 4007â4016. doi:10.1145/2556288.2557142 [50] Vivian Lai, Chacha Chen, Alison Smith-Renner, Q. Vera Liao, and Chenhao Tan. 2023. Towards a Science of Human-AI Decision Making: An Overview of Design Space in Empirical Human-Subject Studies. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (Chicago, IL, USA) (FAccT â23). Association for Computing Machinery, New York, NY, USA, 1369â1385. doi:10.1145/3593013.3594087 [51] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. 2024. Gener- ative AI meets 3D: A Survey on Text-to-3D in AIGC Era. arXiv:2305.06131 [cs.CV] https://arxiv.org/abs/2305.06131 [52] Jenny T. Liang, Melissa Lin, Nikitha Rao, and Brad A. Myers. 2025. Prompts Are Programs Too! Understanding How Developers Build Software Containing Prompts. arXiv:2409.12447 [cs.SE] https://arxiv.org/abs/2409.12447 [53] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. 2023. Magic3D: High-Resolution Text-to-3D Content Creation. arXiv:2211.10440 [cs.CV] https: //arxiv.org/abs/2211.10440 [54] Vivian Liu, Jo Vermeulen, George Fitzmaurice, and Justin Matejka. 2023. 3DALL- E: Integrating Text-to-Image AI in 3D Design Workflows. In Proceedings of the 2023 ACM Designing Interactive Systems Conference (Pittsburgh, PA, USA) (DIS â23). Association for Computing Machinery, New York, NY, USA, 1955â1977. doi:10.1145/3563657.3596098 [55] Do Xuan Long, Duy Dinh, Ngoc-Hai Nguyen, Kenji Kawaguchi, Nancy F. Chen, Shafiq Joty, and Min-Yen Kan. 2025. What Makes a Good Natural Language Prompt? arXiv:2506.06950 [cs.CL] https://arxiv.org/abs/2506.06950 [56]Wendy E. Mackay and Joanna McGrenere. 2025. Comparative Structured Obser- vation. ACM Trans. Comput.-Hum. Interact. 32, 2, Article 14 (April 2025), 27 pages. doi:10.1145/3711838 [57]Atefeh Mahdavi Goloujeh, Anne Sullivan, and Brian Magerko. 2024. Is It AI or Is It Me? Understanding Usersâ Prompt Journey with Text-to-Image Generative AI Tools. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI â24). Association for Computing Machinery, New York, NY, USA, Article 183, 13 pages. doi:10.1145/3613904.3642861 [58]Shareen Mahmud, Jessalyn Alvina, Parmit K. Chilana, Andrea Bunt, and Joanna McGrenere. 2020. Learning Through Exploration: How Children, Adults, and Older Adults Interact with a New Feature-Rich Application. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI â20). Association for Computing Machinery, New York, NY, USA, 1â14. doi:10.1145/3313831.3376414 [59]Damien Masson, Jo Vermeulen, George Fitzmaurice, and Justin Matejka. 2022. Supercharging Trial-and-Error for Learning Complex Software Applications. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA) (CHI â22). Association for Computing Machinery, New York, NY, USA, Article 381, 13 pages. doi:10.1145/3491102.3501895 [60]John Meluso, Amanda Casari, Katie McLaughlin, and Milo Z. Trujillo. 2025. Invisible Labor in Open Source Software Ecosystems. arXiv:2401.06889 [cs.SE] https://arxiv.org/abs/2401.06889 [61] Meshy LLC. [n. d.]. Meshy AI: AI-Powered 3D Model Generator. https://w. meshy.ai/ [62]Steven Moore, Q. Vera Liao, and Hariharan Subramonyam. 2023. fAIlureNotes: Supporting Designers in Understanding the Limits of AI Models for Computer Vision Tasks. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI â23). Association for Computing Machinery, New York, NY, USA, Article 10, 19 pages. doi:10.1145/3544548.3581242 [63]Donald A. Norman. 2013. The Design of Everyday Things: Revised and Expanded Edition. Basic Books, New York, NY. [64] OpenAI. 2024. OpenAI. https://w.openai.com. [65]OpenAI. 2024. Sora: A Large-Scale Model for Realistic Video Generation. https: //openai.com/research/sora. [66]Jonas Oppenlaender, Rhema Linder, and Johanna Silvennoinen. 2024. Prompt- ing AI Art: An Investigation into the Creative Skill of Prompt Engineering. arXiv:2303.13534 [cs.HC] https://arxiv.org/abs/2303.13534 [67]Angela Patrick and Andy McGurgan. 1993. One proven methodology for de- signing robust online help systems. In Proceedings of the 11th Annual Interna- tional Conference on Systems Documentation (Waterloo, Ontario, Canada) (SIG- DOC â93). Association for Computing Machinery, New York, NY, USA, 223â232. doi:10.1145/166025.166084 [68]Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer. 2023. The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arXiv:2302.06590 [cs.SE] https://arxiv.org/abs/2302.06590 [69]Vijay Pereira, Elias Hadjielias, Michael Christofi, and Demetris Vrontis. 2023. A systematic literature review on the impact of artificial intelligence on workplace outcomes: A multi-process perspective. Human Resource Management Review 33, 1 (2023), 100857. doi:10.1016/j.hrmr.2021.100857 [70] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. 2022. DreamFusion: Text-to-3D using 2D Diffusion. arXiv:2209.14988 [cs.CV] https://arxiv.org/abs/ 2209.14988 [71]Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Rad- ford, Mark Chen, and Ilya Sutskever. 2021. Zero-Shot Text-to-Image Generation. arXiv:2102.12092 [cs.CV] https://arxiv.org/abs/2102.12092 [72]John Rieman. 1996. A field study of exploratory learning strategies. ACM Trans. Comput.-Hum. Interact. 3, 3 (Sept. 1996), 189â218. doi:10.1145/234526.234527 [73] Martin P. Robillard and Robert Deline. 2011. A field study of API learning obstacles. Empirical Softw. Engg. 16, 6 (Dec. 2011), 703â732. doi:10.1007/s10664- 010-9150-8 [74]Runway Research. 2024. Runway Gen-2: AI System for Text-to-Video and Video- to-Video Generation. https://runwayml.com/research. [75]Vincent Sitzmann, Michael ZollhĂśfer, and Gordon Wetzstein. 2020. Scene Repre- sentation Networks: Continuous 3D-Structure-Aware Neural Scene Representa- tions. arXiv:1906.01618 [cs.CV] https://arxiv.org/abs/1906.01618 [76]Spline, Inc. [n. d.]. Spline AI: AI-Driven 3D Generation on the Spline Platform. https://spline.design/. [77] Stability AI. 2024. Stability AI. https://stability.ai. [78] Unity Technologies. 2025. Unity â Create and grow real-time 3D games, apps, and experiences for entertainment, film, automotive, architecture, and more. https://unity.com/. [79]Michelle Vaccaro, Abdullah Almaatouq, and Thomas Malone. 2024. When com- binations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour 8, 12 (Oct. 2024), 2293â2303. doi:10.1038/s41562-024- 02024-1 [80] Allison Woodruff, Renee Shelby, Patrick Gage Kelley, Steven Rousso-Schindler, Jamila Smith-Loud, and Lauren Wilcox. 2024. How Knowledge Workers Think Generative AI Will (Not) Transform Their Industries. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI â24). Association for Computing Machinery, New York, NY, USA, Article 641, 26 pages. doi:10.1145/3613904.3642700 [81]H. Peyton Young. 2009. Learning by trial and error. Games and Economic Behavior 65, 2 (2009), 626â643. doi:10.1016/j.geb.2008.02.011 [82]Bhada Yun, Dana Feng, Ace S. Chen, Afshin Nikzad, and Niloufar Salehi. 2025. Generative AI in Knowledge Work: Design Implications for Data Navigation and Decision-Making. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI â25). ACM, 1â19. doi:10.1145/3706598.3713337 [83] J.D. Zamfirescu-Pereira, Richmond Y. Wong, Bjoern Hartmann, and Qian Yang. 2023. Why Johnny Canât Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (Hamburg, Germany) (CHI â23). Association for Computing Machinery, New York, NY, USA, Article 437, 21 pages. doi:10.1145/3544548. 3581388 [84]Beiqi Zhang, Peng Liang, Xiyu Zhou, Aakash Ahmad, and Muhammad Waseem. 2023. Practices and Challenges of Using GitHub Copilot: An Empirical Study. In Proceedings of the 35th International Conference on Software Engineering and Knowledge Engineering (SEKE2023, Vol. 2023). KSI Research Inc., 124â129. doi:10. 18293/seke2023-077 [85]Haoxiang Zhang, Shaowei Wang, Tse-Hsun Chen, Ying Zou, and Ahmed E. Hassan. 2021. An Empirical Study of Obsolete Answers on Stack Overflow. IEEE Transactions on Software Engineering 47, 4 (April 2021), 850â862. doi:10.1109/tse. 2019.2906315 [86]Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-Following Evaluation for Large Language Models. arXiv:2311.07911 [cs.CL] https://arxiv.org/abs/2311.07911