Paper deep dive

From Vibe to Code -- and Back: Lexical Oscillation in the Formation of Design Intent with Generative AI

Daisaku Sato

Year: 2026Venue: arXiv preprintArea: cs.HCType: PreprintEmbeddings: 90

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Generative AI design tools make natural-language prompts a starting point for design, placing new articulation demands on designers. Rather than treating prompts as the transmission of pre-existing design intent, we ask how design intent is formed through situated interaction with AI. Five expert UI/UX designers (11-20 years' experience, M = 15.4) designed landing-page hero sections with a generative AI tool, recorded through think-aloud and retrospective interviews. Using reflexive thematic analysis, we used lexical granularity (L1 vibe, L2 design-domain, L3 operational language) as a sensitizing lens. Rather than moving from vibe to code unidirectionally, designers showed lexical oscillation, including returns from operational specificity to ambiguity. Mismatches with AI outputs were taken up as occasions for designers to reconsider what they meant, and engagement shifted from instruction to consultation. One non-oscillating trajectory -- a negative case -- suggested conceptual misalignment as a tentative boundary for future examination. We position AI as a non-neutral generative interlocutor and ambiguity as a resource for design judgment.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

88,310 characters extracted from source content.

Expand or collapse full text

From Vibe to Code — and Back: Lexical Oscillation in the Formation of Design Intent with Generative AI Daisaku Sato Graduate School of Knowledge Science, Japan Advanced Institute of Science and Technology (JAIST) / Sumitomo Mitsui Finance and Leasing Co., Ltd. Corresponding author: dice310@gmail.com Abstract Generative AI design tools make natural-language prompts a starting point for design, placing new articulation demands on designers. Rather than treating prompts as the transmission of pre-existing design intent, we ask how design intent is formed through situated interaction with AI. Five expert UI/UX designers (11–20 years’ experience, M = 15.4) designed landing-page hero sections with a generative AI tool, recorded through think-aloud and retrospective interviews. Using reflexive thematic analysis, we used lexical granularity (L1 vibe, L2 design- domain, L3 operational language) as a sensitizing lens. Rather than moving from vibe to code unidirectionally, designers showed lexical oscillation, including returns from operational specificity to ambiguity. Mismatches with AI outputs were taken up as occasions for designers to reconsider what they meant, and engagement shifted from instruction to consultation. One non-oscillating trajectory—a negative case—suggested conceptual misalignment as a tentative boundary for future examination. We position AI as a non-neutral generative interlocutor and ambiguity as a resource for design judgment. Keywords: human–AI interaction; generative AI; design intent formation; lexical oscillation; ambiguity as a resource; reflexive thematic analysis 1. Introduction 1.1 Background: Prompt-Driven Design and Emerging Literacy Demands Generative AI tools have rapidly permeated design practice. Tools such as Figma AI, Replit AI, and Cursor now generate UI components directly from natural language text, fundamentally transforming how designers work. Whereas designers previously gave form to mental images through sketches and wireframes, these tools position natural language “prompts” as the starting point for design. This approach gained wide recognition when Karpathy (2025) coined the term “vibe coding” on social media. This phenomenon—in which programmers and designers convey vague impressions (“vibes”) instead of specifying code or design details, delegating concretization to AI—can be situated within the scholarly discourse on end-user programming and natural language programming (Nardi, 1993; Lieberman et al., 2006). The implications are far-reaching: it represents a fundamental paradigm shift in the human–machine relationship. PREPRINT — This paper has been submitted to the International Journal of Human-Computer Interaction (IJHCI) and is currently under review. Scope of terminology: In this study, “vibe coding (vibe to code)” is employed not as a transplantation of a programming technique from the software development context, but as an analytical concept describing the verbalization process by which designers move from vague impressions (vibe) to operational specifications (code). Accordingly, the findings of this study are not directly applicable to programming practice. Under the previous paradigm (deterministic intent negotiation), humans fully defined the “correct answer” and machines faithfully executed instructions. Deviations from instructions constituted “errors”—noise to be eliminated. By contrast, collaboration with generative AI (probabilistic intent negotiation) involves working with outputs generated under uncertainty —outputs shaped by model behavior, interface constraints, and embedded design norms—that may diverge substantially from designers’ emerging expectations. Crucially, such deviations are not necessarily “failures.” Rather, they can function as “cues” for recognizing the ambiguity of one’s own instructions and refining verbalization. 1.2 Problem Statement: From Hidden Intent to Intent Formation in Interaction Polanyi’s (1966) observation that “we can know more than we can tell” captures a central problem in design expertise: much of what designers know cannot be transparently translated into words without changing its form. In this study, we therefore treat articulation not as the extraction of hidden content, but as a situated practice through which what can be noticed, judged, and acted upon becomes progressively shaped. Schön’s (1983) reflection-in-action and Cross’s (2004) account of designerly ways of knowing likewise suggest that design expertise is enacted in practice rather than stored as explicit propositions waiting to be retrieved. This reframes the problem posed by generative AI design tools. A common account of prompt- mediated design asks how designers externalize pre-existing intentions into prompts, as if clearer language simply transmitted already-formed intent. By contrast, this study asks how design intent becomes formed and reconfigured through interaction. Work on ambiguity as a resource for design (Gaver et al., 2003), situated action (Suchman, 1987), and entanglement- oriented accounts of agency (Barad, 2007) suggests that intent is not simply issued from an inner source, but configured through unfolding encounters among people, tools, materials, and situations. Generative AI sharpens this question because its outputs are not neutral reflections of designers’ intentions. A generative AI system does not transparently render an instruction; its outputs may be shaped by training data, interface constraints, and embedded design norms, and these outputs can in turn shape what designers come to mean and intend. Existing research has largely examined prompt-skill acquisition, collaboration efficiency, and output quality. Less is known about the micro-level process through which design intent is formed: the hesitations, mismatches, rephrasings, and realizations that occur as designers move between vague and precise language during interaction with generative AI. This study addresses that gap. 1.3 Research Objectives and Questions This study provides a microscopic qualitative account of how expert UI/UX designers form and reconfigure design intent through iterative interaction with generative AI. As an exploratory study, it does not seek statistical generalization; rather, it aims to develop an empirically grounded analytic vocabulary for describing prompt-mediated design practice. The study makes three contributions. First, it empirically describes lexical oscillation in expert designers’ interaction with generative AI, including episodes in which designers returned from operational specificity to more ambiguous language. Second, it conceptually reframes prompt- mediated design as a process of design intent formation and treats AI as a non-neutral generative interlocutor whose outputs participate in the situated reconfiguration of intent. Third, it positions L1–L3 lexical granularity as a sensitizing analytic lens within reflexive thematic analysis and derives design implications for supporting intent formation in human– AI design workflows. Main Research Question (MRQ): How do expert UI/UX designers form and reconfigure design intent through situated interaction with generative AI, particularly through oscillation across levels of lexical granularity? SRQ1: How are semantic mismatches between generative AI outputs and designers’ emerging expectations associated with the formation and redefinition of design intent? SRQ2: How do designers shift their mode of engagement with generative AI from instruction to consultation, and how is this shift associated with the unfolding formation of design intent? This paper is organized as follows. Section 2 reviews related work on generative AI in design, ambiguity in design practice, situated human–AI interaction, and lexical granularity in prompt-mediated design. Section 3 introduces lexical oscillation and the ECRT schema as sensitizing analytic resources rather than universal models. Section 4 describes the research design, participants, data collection, and reflexive thematic analysis. Section 5 presents the findings, and Section 6 discusses theoretical and practical implications. Section 7 concludes the paper. 2. Related Work 2.1 Generative AI in Design and Prototyping The application of generative AI to the design domain is advancing rapidly. Koch et al. (2019) demonstrated the potential of positioning AI as a creative partner, while Yang et al. (2020) organized the challenges of AI utilization in UX design. A common characteristic of these tools is that users describe “what they want to create” in natural language, and AI interprets that intent to generate visual outputs. However, this “description of intent” itself poses a new cognitive challenge for designers. Limitations of existing research: Most research on generative AI design tools has focused on output quality evaluation and usability assessment. Yet the micro-level cognitive processes of “how designers verbalize” and “what happens during the verbalization process” have not been sufficiently elucidated. 2.2 Human-AI Co-Creation and Prompting Strategies Knoth et al. (2024) demonstrated that prompt creation is a learnable skill dependent on AI literacy. Zamfirescu-Pereira et al. (2023) analyzed difficulties non-experts face in dialogue with LLMs and identified the fundamental challenge of “not knowing what to ask.” Tankelevitch et al. (2024) systematically organized the metacognitive demands of generative AI use. Lee et al. (2025) showed, in a survey of 319 knowledge workers, that high trust in generative AI was associated with reduced critical thinking, while high task-specific self-efficacy was associated with enhanced critical thinking. Sidra and Mason (2025) developed and validated a “Collaborative AI Metacognition” scale, presenting important findings from the perspective of Kahneman’s (2011) dual-process theory. Simkute et al. (2025), writing in the International Journal of Human–Computer Interaction, identified the “ironies of generative AI.” They analyzed the challenges of productivity decline due to user role shifts from “production” to “evaluation,” inefficient workflow restructuring, and interruption, drawing on decades of human factors research. Their macro-level analysis provides a complementary perspective to the micro-level investigation in the present study. Limitations of existing research: Most prompting research has focused on the “structure” and “techniques” of prompts. What remains underexplored is how expert designers form, test, and reconfigure design intent through prompt-mediated interaction, and what cognitive shifts occur during that process. 2.3 Prior Models of Knowledge Conversion and Their Limits for Human–AI Interaction Polanyi’s (1966) concept of tacit knowledge occupies a central position in design research: much of what designers know is enacted in practice rather than held as explicit propositions waiting to be retrieved. Earlier models of knowledge conversion—such as Nonaka and Takeuchi’s (1995) SECI account of how tacit knowledge becomes articulable through dialogue, metaphor, and analogy—were developed on the assumption of human-to-human interaction. They therefore say little about what happens when the “other” in the dialogue is a generative AI system whose outputs are not transparent renderings of intent. We take this as a stepping- stone rather than a framework to apply: rather than asking how a designer’s pre-existing tacit knowledge is transferred to an interlocutor, this study asks how design intent is formed and reconfigured through situated interaction with a non-neutral generative system (see Sections 1.2 and 3). 2.4 From Deterministic to Probabilistic End-User Programming The “vibe coding” phenomenon observed in this study can be theoretically positioned as a paradigm shift in end-user programming (EUP). Generative AI introduces a probabilistic EUP paradigm in which users function as “supervisors” or “curators,” providing high-level intent (Vibe) while delegating logical implementation to the model’s probabilistic nature. This shifts the core competency from “logic description” to “intent negotiation.” 2.5 Research Gap and Positioning of This Study The above review identifies the following gaps: (1) the micro-level process through which design intent is formed during interaction with generative AI has not been systematically traced; (2) prior accounts of knowledge conversion assume human-to-human dialogue and have not been reconsidered for interaction with a non-neutral generative system; and (3) the lexical strategies through which expert designers move between vague and precise language— including returns to ambiguity as a resource—have not been described. Literature collection method: The literature for this review was collected using ACM Digital Library, Scopus, and Web of Science with the search string “generative AI” AND (“design” OR “tacit knowledge” OR “prompt*“) AND (”think-aloud” OR “articulation” OR “design intent”) for English-language publications from 2019–2025. Forward and backward citation tracking (snowballing) and foundational works in cognitive science and design research (Polanyi, 1966; Schön, 1983; Cross, 2004; Gaver et al., 2003) were manually supplemented. This review is a narrative rather than a systematic review, aimed at covering the conceptual scope necessary for the theoretical framing. 3. Theoretical Framework 3.1 Analytic Framework: The ECRT Schema and Lexical Granularity We approach designers’ interaction with generative AI through two sensitizing analytic resources rather than predictive models: the ECRT schema and the L1–L3 lexical-granularity lens. Neither is offered as a universal cognitive model; both serve to orient attention to how design intent is formed and reconfigured in interaction. Lexical oscillation, introduced in Section 3.2, names the observed movement across levels of lexical granularity, including returns from operational specificity toward more ambiguous language. The ECRT schema distinguishes four analytic phases through which a designer’s engagement with an AI output can be traced: Expectation (E), Collision (C), Reflection (R), and Transformation (T). These phases are analytic distinctions, not a mandatory sequence. Rather than positing a fixed cognitive process, the schema foregrounds moments at which a mismatch between a designer’s expectation and a non-neutral AI output becomes an occasion for reconsidering intent. We draw on Schön’s (1983) reflection-in-action to characterize the reflective moment, and on accounts of ambiguity as a resource for design (Gaver et al., 2003) to characterize designers’ returns to less specific language. We do not treat the schema as operationalizing the conversion of pre-existing knowledge. Figure 1. The ECRT schema as a sensitizing analytic resource for tracing moments of expectation, mismatch, reflection, and transformation in the formation of design intent during interaction with non-neutral generative AI outputs. Expectation (E) refers to the designer’s emerging anticipation of what the AI output might or should become; we distinguish vaguer (E-vague) and more specific (E-specific) anticipations as analytic shadings rather than fixed types. Collision (C) refers to a perceived mismatch between the AI output and the designer’s emerging expectation. We retain “collision” as an analytic label within the schema but use “mismatch” in the surrounding analysis to avoid implying a deterministic error signal; we distinguish mismatches concerning meaning, atmosphere, or tone (C-semantic) from those concerning structure, layout, or technical detail (C-syntactic). Reflection (R) refers to moments in which designers reconsider the output, their own wording, or the direction of the design; we note where reflection turned toward the designer’s own way of meaning (R-deep) as distinct from surface dissatisfaction with the output (R- surface). Transformation (T) refers to subsequent changes in the designer’s language, prompt structure, or interactional stance—lexical, structural, or meta-level, such as a shift from instruction-oriented to consultation-oriented prompting. 3.1.1 Relating the Schema to Existing Accounts of Reflection The ECRT schema shares features with established accounts of reflection and learning, but these are different kinds of theory and are not directly commensurable. Kolb’s (1984) experiential learning cycle and Boud et al.’s (1985) account of reflection address learning, whereas Schön’s (1983) reflection-in-action addresses professional practice. We therefore relate the schema to them discursively rather than tabulating them as equivalents. Our use of “collision” is closest to Schön’s notion of a situation that talks back. The difference is not that the schema supersedes Schön’s account, but that the present setting involves a non- neutral generative system whose outputs are not transparent renderings of intent. The same prompt may produce variable outputs, and these outputs may be shaped by training data, interface constraints, and embedded design norms. We therefore use the ECRT schema only to orient the analysis of how designers responded to mismatches in this specific human–AI design setting. The schema is not universal. It is least applicable where AI outputs match expectations and no mismatch arises (for example, with deterministic tools or fully compliant output), where a designer has little habit of metacognitive reflection, and where designer and system do not share a basic working concept of the task. The present data are also limited to Japanese- language interaction, so the applicability of the L1–L3 lens to substantially different linguistic structures is unverified. We treat these as boundaries of the schema’s usefulness rather than as falsification conditions, consistent with its role as a sensitizing resource. 3.2 Lexical Granularity and Lexical Oscillation We classify the vocabulary designers direct at AI into three levels of lexical granularity: L1 (vibe)—impressionistic language such as “modern feel” or “sense of trust”; L2 (design-domain vocabulary)—professional terms such as “hero section” or “minimalist layout,” without operational numerical specification; and L3 (operational)—concrete specifications such as “#000000” or “16px padding.” We use lexical oscillation as the umbrella term for movement across these levels, including returns from operational specificity (L3) toward more ambiguous language (L1/L2). Where designers explicitly articulated an intention behind such a return, we describe these episodes more narrowly as strategic oscillation. Elsewhere, we describe the movement as situated or improvisational, recognizing that not all returns to ambiguity are deliberately planned or verbally articulated as intentional. 3.2.1 Relating Lexical Oscillation to Existing Design Concepts Lexical oscillation is not proposed as a replacement for established design concepts such as iteration, reflection-in-action, divergent–convergent movement, or problem–solution co- evolution (Dorst & Cross, 2001). Rather, it provides a more specific vocabulary for describing movements in prompt language during interaction with generative AI. Iteration can occur at the same lexical level—fine-tuning an operational value (for example, changing a color from #2563EB to #1D4ED8) is iteration, not oscillation. Lexical oscillation, by contrast, concerns movement across levels of granularity, especially returns from operational specificity toward more ambiguous language. Similarly, lexical oscillation can coexist with reflection-in-action, but the two are not identical: reflection is an internal cognitive process, whereas oscillation is an observable movement in the granularity of prompt language. Because this study uses reflexive thematic analysis, these distinctions are not treated as fixed identification criteria. They are analytic distinctions that helped us describe contrasting trajectories in the data, including episodes of return to ambiguity and the negative case in which such return was not productively sustained. We note one limitation directly: our data can only register returns to ambiguity that surfaced in talk or behavior, so movements that remained tacit are likely under-represented. 3.3 Methodological Positioning of Confidence-Satisfaction Scores: Structured Elicitation Device In this study, the 0–10 scores for confidence (“degree to which the prompt conveyed intent”) and satisfaction (“satisfaction with output”) are not used as dependent variables for statistical analysis. With N = 5 and 4–12 units per participant, the statistical pattern-detection power of the scores is insufficient, and such use would be inappropriate. The function of the scores in this study is as a structured elicitation device complementing Ericsson and Simon’s (1993) think-aloud method. The act of score assignment prompts participants to verbalize “why that number,” thereby articulating implicit evaluation criteria as verbal data. Accordingly: Primary data: Verbal content during score assignment (e.g., “I gave it a 5 because the atmosphere is close but the layout is completely different”) Contextual information: The direction and magnitude of score changes are reported as descriptive indicators for contextualizing the intensity and content of verbalizations (e.g., “Accompanying a large negative gap from confidence 8 to satisfaction 1, P005 verbalized the following...”) Not used: Mean scores, variance, inter-participant statistical comparisons, correlation analyses All score reporting in this study is based on this positioning as a “structured elicitation device.” Score changes reported in the results section (Section 5) serve to contextualize verbal data and do not rely on numerical patterns per se. 3.4 Sensitizing Concepts Used in the Analysis Table 1 summarizes the working descriptions of the sensitizing concepts used in the analysis. These descriptions were not treated as fixed coding rules or measurement criteria; rather, they provided provisional analytic guidance and were refined through recursive engagement with the data. Table 1. Analytic Descriptions of Sensitizing Concepts Analytic conceptWorking description Illustrative indicatorsBoundary notes L1 (vibe)Sensory, impressionistic language without design-domain terminology “modern feel,” “sense of trust,” “a ○ atmosphere” Contains UI terminology (then L2); contains color codes or px values (then L3) L2 (design-domain)Design-domain professional vocabulary without operational numerical specification “hero section,” “CTA button,” “grid layout,” “white space” Sensory expression only (L1); contains CSS values or color codes (L3) L3 (operational)Implementable concrete specifications: numbers, code, “#2563EB,” “font- size: 16px,” “padding: 24px” Concept name only (“rounded button” = L2); no numerical values (L2 or below) Analytic conceptWorking description Illustrative indicatorsBoundary notes precise colors Semantic mismatch (C-semantic) Output’s meaning, atmosphere, or tone diverges from expectation “completely different atmosphere,” “the tone is off” Dissatisfaction with structure/layout only (then C- syntactic) Syntactic mismatch (C-syntactic) Output’s structure, layout, or technical detail diverges from expectation “the layout is different,” “it’s not two columns” Dissatisfaction with atmosphere/tone (then C-semantic) Deeper reflection (R- deep) Reconsideration of one’s own way of meaning or premises “my phrasing was ambiguous,” “what was I trying to convey?” Reference to output features only (R- surface); modification instructions only (T) Conceptual alignment Designer and AI share a basic working concept of the task same concept and terminology, output within conceptual range different concepts under the same term (e.g., P005’s “LP”); operational fixes keep failing Lexical oscillationMovement across granularity levels, including returns from L3 toward L1/L2 a return to “let me restart from the image” after specifying numbers same-level edits are iteration, not oscillation; tacit returns are under- represented Note. The distinction between C-semantic and C-syntactic was not theorized a priori but emerged abductively from P001’s data (Section 4.5). Conceptual alignment was developed as a boundary notion through analysis of P005’s disengagement. 4. Methodology 4.1 Research Design Epistemological stance: This study draws on an interpretivist epistemology and adopts a constructivist ontology (Creswell & Poth, 2018). Designers’ cognitive processes are understood as socially and contextually constructed; we do not assume a single objective reality measurable independently of context. Understanding participants’ unique meaning-making and its generative processes thus aligns with the research objective. This epistemological stance motivates the selection of abductive analysis, first-person data collection through think-aloud methods, and interpretive coding. This study was designed as an exploratory qualitative inquiry (Creswell & Poth, 2018), based on an abductive reasoning approach. The initial ECRT schema served as a sensitizing starting point and was refined through recursive engagement with the data. This iterative oscillation between theory and data is characteristic of the abductive approach (Tavory & Timmermans, 2014) and is distinguished from purely deductive hypothesis testing and purely inductive grounded theory. Think-aloud protocols (Ericsson & Simon, 1993) were employed for data collection. Researcher positionality and reflexivity: The author is an insider researcher with 25 years of UI/UX design practice, sharing a professional domain with the participants. Consistent with reflexive thematic analysis (Braun & Clarke, 2019), we treat this position as an analytic resource rather than solely as a source of bias to be eliminated: domain familiarity sharpened sensitivity to designers’ implicit verbalization patterns and to the nuances of professional terminology. At the same time, we remained reflexively attentive to the risk that the author’s own verbalization habits could operate as implicit interpretive standards. We addressed this reflexively by (a) maintaining a reflexive journal documenting how the author’s assumptions shaped coding and theme development; (b) engaging a critical friend, a doctoral researcher in HCI, who reviewed selected coded extracts, challenged candidate themes, and discussed discrepant interpretations with the author; and (c) actively seeking discrepant and negative cases, most notably P005’s disengagement, as occasions to revise rather than confirm the developing account. 4.2 Participants Selection criteria included: (1) 10+ years of UI/UX design practice experience, (2) lead designer experience on commercial products, and (3) 6+ months of professional use of generative AI tools. Table 2. Participant Demographics ID Years Exp.DomainAI Exp.AI Freq.TypeOutcome P00115 yearsUI/UX, SaaS12 months DailyLeapfrogSuccessful (sat. 7) P00213 yearsUI/UX, Healthcare8 months3–4 times/week RecoveryRecovery (0→10) P00318 yearsUI/UX, FinTech10 months 3–4 times/week Problem-solvingRecovery (2→9) P00420 yearsUI/UX, Manufacturing SaaS 12 months DailyGradualGradual (3→6) P00511 yearsUI/UX, Enterprise6 months2–3 times/month Disengagement ★ Disengagement (sat. 1×2) Note. P005 is the negative case (deviant case). = confound candidate variable (P005 shows★ the lowest value and confounding with session behavior cannot be ruled out). AI use history = months of professional use of generative AI tools. AI use frequency = self-reported typical frequency confirmed in retrospective interviews. Satisfaction = final score (0–10, used as structured elicitation device, see Section 3.3). P005 showed the lowest values on three variables: years of experience (11 years), AI use history (6 months), and AI use frequency (2–3 times/month), and the possibility that these background differences influenced session behavior cannot be excluded (see Section 6.6.1). Table 3. Deviant Case Analysis: Contrast Between P005 and Successful Participants (P001–P004) Comparison DimensionP001–P004 (Successful)P005 (Disengagement) AI experience8–12 months (M = 10.5)6 months (minimum threshold) AI use frequency3–4 times/week to daily2–3 times/month (lowest) Design experience13–20 years (M = 16.5)11 years (lowest) Strategy shift observationPresent (all 4 participants)Absent (linear elaboration only) Consultation-mode promptsSpontaneously adoptedNot observed Failure attribution styleInternal: “my verbalization”External: “AI’s capability” AI expectation setting“Draft generator” / “starting point” “It should understand me” Conceptual alignmentAchieved in early sessionNot achieved (“LP” mismatch) Maximum negative gap-5 (P002, recovered)-7 (no recovery) Note. P005’s AI experience of 6 months (minimum selection criterion), AI use frequency (2–3 times/month, lowest among participants), and absence of strategy shift suggest the possibility that conceptual alignment requires a threshold of AI interaction experience (including not only use duration but also use density) (tentative proposition). Gap values descriptively indicate the direction of score changes as structured elicitation devices (see Section 3.3). 4.2.1 Sample Size Rationale The sample of N = 5 was strategically designed to maximize Information Power for theory building (Malterud et al., 2016). Following their five-dimension framework: (a) study aim— narrow and specific (micro-level cognitive processes during AI dialogue); (b) sample specificity —participants possessed an average of 15.4 years of professional design experience, constituting a “high-signal” group with dense, relevant knowledge; (c) analytic focus—the L1– L3 lexical-granularity lens, together with prior work on reflection-in-action and ambiguity as a resource, focused the analysis, while the ECRT schema is treated as a sensitizing schema rather than as established theory; (d) quality of dialogue—the think-aloud protocol combined with retrospective interviews and structured elicitation yielded rich verbal data per participant; and (e) analysis strategy—abductive case-oriented analysis guided by sensitizing-concept descriptions and an audit trail. This combination yields high information power per participant. The sampling strategy also successfully captured a theoretically significant negative case (P005). Limitation: The “disengagement” pattern is represented by P005 alone, and the breadth and variability of this subcategory could not be established from the present data. Conclusions regarding the disengagement pattern are positioned as tentative propositions. 4.2.2 Trustworthiness Following Lincoln and Guba (1985), we addressed trustworthiness rather than reliability. Credibility was supported by think-aloud and retrospective interview data, repeated engagement with each transcript and screen recording, critical-friend dialogue, and negative- case analysis (P005). Transferability was supported by thick description of the participants, task, tool, and interaction setting (Sections 4.2–4.4). Dependability and confirmability were supported by an audit trail documenting unit segmentation, the evolution of the sensitizing concepts, analytic memos, and theme-development decisions, together with a reflexive journal recording how the author’s position shaped interpretation. 4.2.3 Ethical Considerations Prior to session commencement, all participants received comprehensive written and oral explanations of the following: (1) the objectives and procedures of the study, (2) data handling methods (including audio recording, screen recording, and anonymization procedures), (3) the voluntary nature of participation and the right to withdraw at any time without providing a reason, and (4) complete deletion of all data in the event of withdrawal. Written informed consent was obtained from each participant before the session began. The consent form explicitly specified that data use would be limited to academic research purposes, described data storage methods (encrypted storage), anonymization procedures (replacement of personally identifiable information with pseudonyms), and post-study data management policies. All participants signed the consent form after confirming their understanding of its contents. The following data protection procedures were implemented. Session verbalizations were audio-recorded with participants’ consent, and audio files were deleted upon completion of transcription. Screen-sharing recordings were used to document the chronological sequence of prompts and outputs, and anonymization processing was performed after analysis completion. Personally identifiable information in transcripts (names, company affiliations, project names, etc.) was replaced with pseudonyms. All data were stored in encrypted storage with access restricted to the research team. The consent form explicitly stated that participants could withdraw at any time without providing a reason, and that all data from withdrawing participants would be completely deleted. No compensation was provided for participation in this study. Ethics approval statement: Formal ethics committee approval was not obtained for this study. The reason is as follows. At the time of data collection (January 2026), the author was affiliated with Sumitomo Mitsui Finance and Leasing Co., Ltd., a private-sector financial services organization that does not maintain an institutional review board (IRB) or ethics committee for non-medical behavioral research. Because no institutional ethics review mechanism was available, the author designed and conducted the study in strict accordance with the ethical principles of the Declaration of Helsinki (World Medical Association, 2013) and the Ethical Principles of Psychologists and Code of Conduct (American Psychological Association, 2017). The study posed minimal risk to participants: they voluntarily performed a routine professional design task (creating a landing page hero section) using commercially available tools in a familiar work setting, with no deception, no collection of sensitive personal data, and no physical or psychological intervention. Written informed consent was obtained from each participant prior to session commencement, documenting: (a) the objectives and procedures of the study, (b) data handling methods including audio and screen recording, (c) anonymization procedures, (d) the voluntary nature of participation, and (e) the right to withdraw at any time without consequence, with complete data deletion guaranteed upon withdrawal. Ethical oversight for subsequent phases of this research program will be provided by the institutional review board of the Japan Advanced Institute of Science and Technology (JAIST). 4.3 Task and Tools Participants designed the hero section of a landing page for a fictional B2B SaaS service. Tool: Figma Make (version 2025.11). The underlying generative model was not disclosed by the vendor. Data collection period: January 15–25, 2026. The hero section task was selected because it requires designers to integrate multiple design dimensions—visual hierarchy, brand tone, typographic choices, and spatial composition— within a single, bounded artifact. This multidimensional nature naturally elicits vocabulary spanning all three granularity levels (L1 abstract/emotional, L2 structural, L3 technical), providing a suitable context for observing the verbalization of implicit design knowledge. Positioning of replicability regarding tool and model opacity: This study used Figma Make (version 2025.11). The underlying generative model was not disclosed by the vendor and may change over time. Therefore, this study defines replicability not at the level of bitwise output replication but at the level of procedural and analytic replicability. Specifically, replicability is assessed by whether the patterns described here—lexical oscillation and the ECRT schema—can be re-observed in similar generative AI environments using the same protocol and sensitizing concepts. To this end, we provide the sensitizing-concept descriptions (Table 1), unit segmentation rules (Section 4.5), and an audit trail (Supplementary Material) to ensure auditability. Table 4. Fixed and Non-Fixed Elements of the Experimental Environment CategoryElementDetails/Notes FixedToolFigma Make version 2025.11 (same version across all sessions) TaskB2B SaaS LP hero section (same brief for all participants) Session duration30-minute AI dialogue (uniform across all participants) Prompt input methodText input only (no voice or image input) UI settingsDefault settings (no custom presets or templates) Data collection periodJanuary 15–25, 2026 (11 days) CategoryElementDetails/Notes LanguageJapanese prompts only Non-fixed (uncontrollable) Underlying generative modelNot disclosed by vendor (Figma Inc.). Model type, version, and update timing not controllable by researchers Temperature parameters, etc.Uncontrollable. Figma Make UI has no model parameter adjustment functionality Probabilistic output variationIdentical prompts may generate different outputs (see mini-verification below) Note. Non-fixed elements constitute the primary threat to this study’s replicability. That said, since the object of analysis is not AI output quality but designers’ cognitive responses to output (ECRT schema), output variation does not directly compromise analytical validity (elaborated in Section 6.6). 4.3.1 Output Variation Verification To evaluate the impact of underlying model non-fixedness on analytical conclusions, a mini- verification was conducted after data collection completion. Three representative prompts used during P001’s session were re-executed three times each using the same tool and settings, and output variation was assessed. Verification results: (1) Layout structure (number of columns, section arrangement) was generally consistent in 3 of 3 trials. (2) Specific color and typographic choices differed in 2–3 of 3 trials. (3) “Atmosphere” and “tone” similarity were evaluated based on the following rubric. Tone similarity rubric: Three evaluation axes—(i) trustworthiness, (i) approachability, (i) modernity—were each rated on a three-level scale (low/medium/high), with “similar” defined as matching the original output on 2 or more of 3 axes. The author and critical friend independently evaluated the results, and in the majority (7 of 9) of 3 prompts × 3 trials = 9 comparisons, both evaluators judged “similar” (agreement: 7/9). This verification is exploratory and does not constitute proof of robustness. The 2 disagreements both involved the “approachability” axis. Implications for conclusions: Probabilistic output variation primarily appeared at the L3 level (concrete specifications: color, font size, etc.), while variation at the L1 level (atmosphere/tone) was relatively small. The central patterns discussed in this study—lexical oscillation, trajectories traced with the ECRT schema, and interactional responses to mismatches—concern designers’ cognitive processes in response to output rather than specific output content, and the possibility of L3-level variation overturning conclusions was not observed within this verification scope. However, this verification is limited and exploratory (n = 9; 3 prompts × 3 trials) and was not designed for systematic evaluation of output variation. We do not claim that output robustness has been demonstrated by this result. Systematic replicability verification remains a task for future research. 4.3.2 Prompt-Output Log Examples (Anonymized) Here, U denotes an interaction unit—one prompt submission, the generated AI output, and the participant’s verbalization during output review; the full segmentation procedure is described in Section 4.5. Below, we present the prompt and output summary for P001’s Unit 3 (in which C-semantic collision → R-deep reflection was observed). Prompt (P001-U3): “Create a hero section for a modern corporate site that conveys trust but isn’t too rigid. I want blue tones and a CTA button that stands out.” AI output summary: A corporate-style layout with a dark navy background and white text. The CTA button was in fluorescent green. The overall impression was close to “an official financial institution website,” diverging from P001’s intended tone of “modern but approachable.” P001’s verbalization (think-aloud): “Hmm, this isn’t right. Too rigid. There’s trust, but it’s not modern... Actually, what did I mean by ‘modern’? Maybe I wasn’t clear about that myself.” Analytic reading: this episode was read as a semantic mismatch (tone), followed by deeper reflection on the ambiguity in the designer’s own verbalization, and then a lexical adjustment (changed to “SaaS startup style” in next prompt) The above is a representative example of C-semantic collision → R-deep reflection. Below, we present an additional lexical oscillation example and P005’s consecutive disengagement unit sequence to ensure representativeness of examples. Example 2. P002-U8→U9 (lexical oscillation example: large-amplitude oscillation) Prompt (P002-U8): “Make it a corporate, trustworthy layout. Keep the navigation simple.” AI output summary: A formal layout based on corporate navy and white. A polished impression, but the tone diverged from the “approachability” P002 had been exploring. P002’s verbalization (think-aloud): “Yeah, this isn’t bad per se... but what was the ‘atmosphere’ of this page supposed to be? I was aiming for something more friendly, but somehow I got pulled toward the rigid direction.” Analytic reading: this episode was read as a semantic mismatch (overall tone divergence from intent), followed by deeper reflection that re-examined the goal itself (“what was...supposed to be?”), and then a meta-level shift (an intentional return from L3 to L1, changing strategy). Example 3. P005-U3→U4→U5 (Disengagement sequence) Prompt (P005-U3): “Make the font a bit bigger, darken the header color a bit, and round the button corners.” AI output summary: Slight font size increase, darker header color. But button corner rounding was not applied, and the overall layout balance differed from P005’s expectations. P005’s verbalization (think-aloud): “Even after explaining this much, it still doesn’t work. Do I have to be even more specific?” Analytic reading: this episode was read as a syntactic mismatch (partial non- implementation), followed by surface-level reflection (dissatisfaction attributed to the AI’s comprehension), and then a lexical adjustment (further detailed instructions were added) Prompt (P005-U4): “Round the button corners with radius 8px. Make the header #2C3E50. Font size: body 16px, heading 24px.” AI output summary: The specified values were largely reflected, but the overall “atmosphere” diverged further from P005’s implicit image. P005’s verbalization (think-aloud): “The numbers are right, but something’s off. I don’t know how to explain it anymore.” Analytic reading: this episode was read as a semantic mismatch (overall impression mismatch despite numerical compliance), followed by surface-level reflection (“don’t know,” an attribution fixation with no reflection on the designer’s own verbalization), and then disengagement (the lowest score was assigned in the next unit, with an expressed intent to end the session) P005’s three-unit sequence illustrates a trajectory in which, despite continued operational detailing (L2→L3 direction), concept-level divergence remained unresolved, and disengagement was reached without consultation-mode shifts or R-deep occurring. The contrast with P001-U3 (Example 1) and P002-U8 (Example 2) illustrates how the success or failure of conceptual alignment co-occurs with subsequent cognitive processes. 4.4 Data Collection Procedure Each session lasted approximately 60 minutes per participant: introduction (5 min), practice time (10 min), image sharing (5 min), AI dialogue session (30 min), and retrospective interview (10 min). Note: No formal pilot study was conducted prior to data collection. The 10-minute practice phase within each session functioned as familiarization with the think-aloud protocol and AI tool, but did not constitute systematic pre-testing of the research protocol, coding approach, or scoring procedures. Future replications should incorporate a dedicated pilot phase to identify potential ambiguities in the coding approach. 4.5 Analytic Approach: Reflexive Thematic Analysis We analyzed the data using reflexive thematic analysis (Braun & Clarke, 2006, 2019), in which coding is understood as an interpretive act shaped by the researcher’s position rather than as the consistent extraction of stable, pre-existing meanings. This approach aligns with the study’s interpretivist and constructivist epistemology (Section 4.1). Unit of analysis. The unit of analysis was one interaction unit consisting of a prompt submission → AI output generation → verbalization during output review. A new prompt submission marked a unit boundary; multiple verbalizations about the same output were consolidated into one unit; the point at which the participant assigned a score marked the unit endpoint. Immediate additions to a preceding prompt (within five seconds) and re-submissions of the same prompt due to system errors were consolidated into the same unit. Practice-session units and units in which no output was generated due to system errors were excluded. (Borderline example: in P002’s Units 7–8, a prompt modification such as “a bit more...” was re- submitted after three seconds and consolidated under the five-second rule.) Of 50 units, 34 were analyzed (P001: 6, P002: 12, P003: 7, P004: 4, P005: 5). Because P002 contributed a larger share of units, we foreground participant-specific readings and representative contrasting trajectories rather than aggregate frequencies. Sensitizing concepts. The L1–L3 lexical granularity scheme and the ECRT schema were used as sensitizing concepts (Blumer, 1954) that oriented our initial engagement with the data; they served as starting points for interpretation, not as measurement instruments. Consistent with an abductive logic (Tavory & Timmermans, 2014), these concepts were held in dialogue with the data and were refined, qualified, and in places restructured through analysis. The distinction between semantic and syntactic mismatch, for instance, was not pre-set but emerged from P001’s contrasting responses to “atmosphere” and “layout” divergences; conceptual alignment was developed as a boundary notion through P005’s disengagement. Analytic process. Following the recursive phases of reflexive thematic analysis (Braun & Clarke, 2006), the author (a) became familiar with the data through repeated engagement with transcripts and screen recordings; (b) generated initial codes; (c) constructed candidate themes; (d) reviewed themes against the data; (e) defined and named themes; and (f) produced the analytic narrative reported in Section 5. Coding and theme development were iterative rather than linear. Critical friend and audit trail. A doctoral researcher in HCI acted as a critical friend (Braun & Clarke, 2019), reviewing selected coded extracts, challenging candidate themes, and discussing discrepant interpretations with the author. The purpose was not to establish convergence or reliability but to prompt reflexivity and surface alternative readings. As a transparency measure, code-level documentation, segmentation decisions, analytic memos, and theme-development records are provided as an audit trail in the Supplementary Material (Tables S1–S3 and the accompanying confusion matrices), not as indices of reliability. 5. Results Score values reported in this section constitute contextual information as structured elicitation devices (Section 3.3) and serve to support interpretation of verbal data. 5.1 Participant Overview Four of five participants (P001–P004) expressed satisfaction during sessions. In contrast, P005 assigned the lowest score (1) twice consecutively and stated “I feel like giving up.” The contrasting lexical granularity trajectories of representative participants (P001 vs. P005) are shown in Figure 2. 5.1.1 Participant-level patterns (audit-trail summary) For transparency, participant-level occurrence patterns of the analytic labels across the 34 units are documented as an audit-trail artifact in the Supplementary Material (Table S2). Consistent with reflexive thematic analysis, we do not treat these counts as findings or as measures of effect; they are provided so that readers can trace how the sensitizing concepts were applied across participants. 5.1.2 Semantic mismatch and deeper reflection (qualitative reading) Across participants, semantic mismatches (divergences of meaning, atmosphere, or tone) tended to be followed by deeper reflection on the designer’s own way of meaning, whereas syntactic mismatches more often prompted surface correction. We read this as a qualitative pattern in the trajectories rather than as a quantified association; the representative recovery trajectory (P002) is presented in Section 5.4 (with a large-amplitude oscillation example, P002 U8→U9, in Section 4.3.2), and the supporting unit-level documentation is provided in the Supplementary Material (Table S3). The clearest exception was P005, for whom semantic mismatch did not turn into reflection on their own wording (Section 5.4). 5.2 Two Contrasting Trajectories Across Lexical Granularity Figure 2 contrasts the lexical granularity trajectories of two representative participants, illustrating the core distinction between lexical oscillation and a non-oscillating trajectory. P001 illustrates the oscillatory pattern: after moving toward operational specificity (L3), P001 returns to more ambiguous, exploratory language (L1)—a return to ambiguity—before moving again toward L3. This return reopened the design space and was associated with productive reconfiguration. P005, by contrast, follows a largely unidirectional move toward L3 without such a return, a non-oscillating trajectory that ended in disengagement. Figure 2. Representative lexical oscillation and a non-oscillating trajectory across L1–L3 lexical granularity. The y-axis represents lexical granularity: L1 vibe, L2 design-domain vocabulary, and L3 operational language. Dotted arrows indicate returns from operational specificity toward ambiguity. P001 illustrates lexical oscillation, whereas P005 illustrates a non-oscillating trajectory ending in disengagement. 5.3 Finding 1: From Instruction to Consultation (Response to SRQ2) Designers did not only refine their instructions; several shifted the mode of their engagement, moving from issuing instructions toward consulting the AI for options while retaining decision authority themselves. This shift widened the space of possibilities a designer considered and, in doing so, contributed to the formation of intent rather than to its mere transmission. The shift took different forms. P001 moved early (at Unit 2 of 6): after a semantic mismatch, P001 verbalized “it’s faster to have several options shown rather than deciding myself” and then prompted “show me three patterns,” redefining the AI’s role from instruction-executor to proposal-generator. P003 reached a similar point through a recognition of their own limits (“I was trying to decide everything myself, but the AI might have more options”), adopting consultation as a reallocation of cognitive resources rather than for speed. P004, the most experienced (20 years), initially resisted (“First, I want to accurately convey what’s in my head”) and only after repeated gaps—realizing that “fine-tuning each time doesn’t change things much”—asked the AI to “suggest two directions.” P004’s case suggests that a control orientation, possibly associated with long experience, may delay the shift, with accumulated mismatches eventually functioning as its trigger. P005 did not make this shift (see Table 3): operational detailing continued without a move to consultation, alongside the disengagement described in Section 5.4. We read the consultation shift as one route through which designers reformed their intent in interaction, while noting that its absence in P005 co-occurred with several features and cannot be attributed to any single cause. 5.4 Finding 2: How Mismatches Became Occasions for Reconsidering Intent (Response to SRQ1) When the mismatch between an AI output and a designer’s expectation concerned tone, atmosphere, or meaning rather than layout or structure, it was often taken up not merely as an occasion for correction but as an occasion for reconsidering what the designer had been trying to mean. The clearest instance of intent forming through interaction came from P001. Reacting to an output whose feel diverged from what they had imagined, P001 asked aloud, “What did I mean by ‘modern’?” This was not the recovery of a fully formed intention waiting to be expressed; it was the moment at which the intention began to take shape. The question made an evaluative criterion, until then held tacitly, available for reconsideration. This is consistent with our framing (Section 1.2): articulating design knowing does not report a pre-existing intention so much as shape it (cf. Polanyi, 1966). Tracking P001’s six units, mismatches of meaning dominated early (“the atmosphere is completely different,” “the tone is off”) and gave way to structural adjustments later (“column width,” “button placement”), suggesting a staged movement in which conceptual footing was settled before operational fine-tuning. P002 showed a recovery trajectory. Reviewing early outputs that felt “completely different from what I was thinking... I couldn’t verbalize it,” P002 reconsidered the goal itself and shifted to consulting the AI for options (“it’s better to have the AI suggest options and I select”), ultimately reaching a satisfying result. P003 exhibited a distinctive comparative reading of mismatches: in several semantic divergences they treated the AI output as an alternative rather than an error (“this is interesting in its own way, but what I wanted to say was...”), re- confirming their own intent while acknowledging the output’s value—a flexibility we relate, tentatively, to their experience without ruling out other factors. P005 is the negative case. Tracking their verbalizations chronologically reveals a consistent external-attribution frame: “It’s different from what I expected,” “Maybe I should write in more detail,” “Even after explaining this much, it still doesn’t work,” “I don’t know how to explain it anymore”—ending in disengagement. In P001–P004, deeper reflection was accompanied by a shift toward treating one’s own verbalization as part of the issue, for which semantic mismatch acted as a trigger; in P005, by contrast, mismatch co-occurred with external attribution, the absence of deeper reflection, and disengagement. We read this contrast as suggesting that a shared working concept may support the productive use of mismatch, while noting that experience differences and other factors cannot be excluded; the boundary is therefore tentative (Section 6.6.1). 5.5 Finding 3: Lexical Oscillation — Returning to Ambiguity as a Resource The central pattern in the data was that lexical granularity did not increase monotonically. Rather than moving steadily from vibe toward code, designers repeatedly returned from operational specificity toward more ambiguous language. We call this movement lexical oscillation. It did not function as regression or error correction but as a way of reopening the design space, and it was associated with eventual convergence or recovery (P001), whereas its absence accompanied disengagement (P005). Where a designer articulated a reason for such a return, we describe the episode as strategic oscillation; where the movement appeared unplanned or was not verbalized as intentional, we describe it as situated or improvisational. We draw this distinction descriptively rather than as a fixed test, and we note its limitation directly: our data can only register returns to ambiguity that surfaced in talk or behavior, so movements that remained tacit are likely under- represented. Several episodes illustrate the pattern. After specifying operational detail (“#2563EB, font-size: 48px, line-height: 1.2”), P001 returned to vibe—“Something with more startup energy”— explaining, “I specified too many numbers; let me restart from the image.” P002, deep in operational detail on a CTA button (“border-radius: 24px, background #F6B35”), paused and returned to vibe: “Wait, what was the ‘atmosphere’ of this page again? I was too deep in the details and lost the overall feel”; the subsequent direction shifted qualitatively from a solid, corporate feel toward a warmer, more approachable one. P003 returned more narrowly, from a settled grid toward “a more calm, professional impression,” keeping the structure and reworking the palette. By contrast, P004’s adjustment from “padding: 32px” to “padding: 24px” stayed at the operational level: this is same-level iteration, not oscillation. P005’s move from “make the font bigger” to “something just... better” was accompanied by “I don’t know what to do anymore”—an expression of confusion rather than an intentional return, and so not an instance of productive oscillation. P004 also showed a distinctive, gradually narrowing pattern across cycles, verbalizing that “going back and forth several times to gradually improve accuracy suits me better”—a controlled exploration underpinned by long experience that contrasts with the more improvisational returns of P001 and P002, indicating that variation exists even among experts. 5.6 Integrated Response to the Main Research Question Taken together, the findings suggest that expert designers did not arrive with fully formed intent that they then put into words; rather, design intent was formed and reconfigured through interaction with the AI’s outputs. The central pattern was lexical oscillation— movement across levels of lexical granularity, including returns from operational specificity toward ambiguity, which kept ambiguity available as a resource for design judgment. Two further movements accompanied this pattern: mismatches of meaning were taken up as occasions for designers to reconsider what they had been trying to mean (SRQ1), and engagement shifted from instruction to consultation (SRQ2). Across these movements, the AI functioned not as a neutral mirror but as a non-neutral generative interlocutor whose outputs participated in the reconfiguration of intent. The one trajectory in which such reconfiguration was not sustained—P005, who continued operational detailing without returning to ambiguity and disengaged—is read here as a negative case, suggesting conceptual misalignment as a tentative boundary rather than a generalizable condition. 6. Discussion 6.1 AI as a Non-Neutral Generative Interlocutor A central question raised by our findings is what kind of partner a generative AI system is in design work. The view we set aside is that of a neutral instrument that transparently renders a designer’s intent—a mirror that simply reflects back what was put in. Our data are not well described by that view: the same prompt could yield different outputs, and outputs frequently returned something the designer had not anticipated. We therefore characterize the AI not as a mirror but as a non-neutral generative interlocutor: a participant whose outputs actively enter into, rather than merely reflect, the formation of design intent. A parallel reconfiguration has been described in software development: Meske et al. (2025) frame vibe coding as a shift in the mediation of developer intent from deterministic instruction toward probabilistic inference, in which intent is formed through sustained human–AI dialogue rather than issued in advance— a dynamic that, although it arises in programming rather than design (cf. Section 1.1), is continuous with the non-neutral, intent-shaping role we attribute to generative AI here. P001’s “What did I mean by ‘modern’?” is instructive here. The output that provoked this question was not a faithful echo of their instruction but a divergent rendering; it was precisely the non-neutrality of the output—its capacity to return something unanticipated—that made it generative for intent formation. Had the system simply mirrored their words, there would have been nothing to reconsider. This connects to the “ironies of generative AI” identified by Simkute et al. (2025) in this journal. The same property that can be disruptive—outputs that diverge from what users expect—can, under some conditions, be productive: divergence that prompts reconsideration rather than mere frustration. Within our data, this productive uptake was most evident where designer and AI shared a basic working concept of the task; where that footing was absent, as with P005, the same divergence was met with frustration and disengagement. We offer this as a tentative boundary rather than a general claim, recognizing that the underlying model was undisclosed and that outputs may be shaped by training data, interface constraints, and embedded design norms. 6.2 Lexical Oscillation and the Sketching Tradition This account connects to earlier design research on sketching and material practice. Buxton (2007) emphasizes that sketching keeps design possibilities provisional and open. Goldschmidt (1991), more pointedly for the present study, explicitly defines the “dialectics of sketching” as an “oscillation of arguments which brings about gradual transformation of images”—a formulation in which external representations do not record pre-existing ideas but help generate them. Vallgårda and Fernaeus (2015) similarly position bricolage as a non- hierarchical negotiation between designer and material. Although these accounts focus on physical and visual practices, the lexical oscillation observed in the present study shares a structural family resemblance: designers shift between vague and operational language not to transmit a fixed intent but to keep design intent provisional, revisable, and responsive to what the generative system produces. The medium differs—sketch lines versus prompt tokens—but the underlying movement, generative rather than transmissive, is recognizably continuous with this earlier tradition. 6.3 The Paradox of Specificity The received wisdom is straightforward: more specific prompts yield better results. Our findings complicate this notion. Expert designers moved between specificity and ambiguity (Figure 2). P005’s case suggests the proposition that “conceptual alignment precedes operational alignment.” This does not mean that specificity is unhelpful. Rather, the findings suggest that specificity becomes useful only after a sufficient working orientation has been established. Operational detail can refine an emerging direction, but when introduced before a shared conceptual footing is in place, it may prematurely narrow the search space or amplify misalignment. The contribution of lexical oscillation is therefore not a rejection of prompt specificity, but a more situated account of when designers move into specificity and when they return to ambiguity. 6.4 Design Patterns and UI Requirements Design Pattern 1: Conceptual Alignment Verification UI. Problem: In the negative case, operational detailing without a shared working concept was associated with disengagement; this suggests, but does not establish, a design opportunity for conceptual-alignment support. Lee et al. (2025) showed that overreliance on generative AI suppresses critical thinking, and conceptual alignment failure may be related to this risk. Solution: An interaction pattern that confirms basic concepts before operational details. Design Pattern 2: Consultation Shift Nudge. Problem: In the observed sessions, participants facing consecutive mismatches sometimes persisted with the same prompting mode rather than switching strategy. Solution: A nudge encouraging a switch to consultation-mode prompting upon detecting consecutive negative gaps. Design Pattern 3: Lexical Granularity Level Feedback. Problem: The observed trajectories suggest that participants may benefit from support in noticing shifts in the granularity of their own prompts. Solution: Real-time feedback of lexical granularity level (L1/L2/L3) during prompt input. Validity assessment of design patterns: The three design patterns above are design hypotheses derived from exploratory data and have not undergone systematic validity assessment prior to implementation. However, we organize the derivation rationale, design requirements, and anticipated failure conditions for each pattern and attempt a tentative validity assessment. DP1 (Conceptual Alignment Verification UI) was motivated by the negative-case trajectory in P005, in which operational detailing without a shared conceptual footing was followed by disengagement (Section 5.4). An anticipated failure condition is that the verification step may interrupt users’ natural workflow and paradoxically promote disengagement. DP2 (Consultation Shift Nudge) was derived from the successful “instruction → consultation” shift patterns observed in P001–P004 (Section 5.3). An anticipated failure condition is that inappropriately timed nudges may violate user autonomy. DP3 (Lexical Granularity Feedback) is based on the hypothesis that consciousness of lexical oscillation supports metacognition. An anticipated failure condition is that feedback may increase cognitive load and particularly inhibit free exploration at the L1 level. Empirical evaluation of these design patterns (e.g., prototype implementation and expert evaluation, usability testing) is positioned as a future research task (see Section 7.3). 6.5 Linguistic and Cultural Considerations All participants in this study were Japanese designers who prompted in Japanese. Japanese interaction in this dataset included frequent subject omission, context-dependent interpretation, and numerous hedging expressions (e.g., “~na kanji de” [with a ~ish feel]). This repertoire of ambiguous expressions may have been associated with the emergence of L1 (Vibe) level prompting. Scope limitation: The association between the above linguistic features and the lexical oscillation observed in this study is merely a tendency observed within the present data scope (N = 5, Japanese speakers only). The causal claim that the high-context nature of Japanese facilitates lexical oscillation cannot be derived from this study design. Whether lexical oscillation is a pattern specific to Japanese speakers or a language-independent expert strategy requires cross-linguistic comparative research (see Section 7.3). Communication norms rooted in high-context culture (Hall, 1976) may be associated with a tendency for designers to expect “intent” to be inferred from minimal cues. P005’s frustration may partially reflect culturally specific expectations regarding the inferential capacity of communication partners. That said, multiple alternative explanations exist for P005’s disengagement beyond linguistic and cultural factors (see Section 6.6 below). 6.6 Limitations and Boundary Conditions 6.6.1 Alternative Explanations for P005’s Disengagement Against this study’s interpretation of P005’s disengagement as “lack of conceptual alignment,” the following alternative explanations exist. Since these cannot be fully discriminated from a single negative case, we organize the refutability of each alternative. Table 5. Analysis of Alternative Explanations for P005’s Disengagement Alternative Explanation Evidence in Present DataRefutabilityJudgment Difference in years of experience P005 had 11 years (lowest). P001-P004 had 13-20 years (M = 16.5) The gap between P003 (18 years) and P005 (11 years) is large, but between P002 (13 years) and P005 (11 years) is small. Years of experience alone do not sufficiently explain disengagement Partially refutable. However, cannot be fully excluded as a confound Difference in AI use experience P005 had 6 months, 2-3 times/month (minimum criteria). P001-P004 had 8-12 months, 3+ times/week Possible threshold effect of AI experience. Whether the difference in use frequency (2-3 times/month vs. 3-4+ times/week) is qualitatively different from use duration (6 vs. 8 months) cannot be determined from the present data Not refutable. Threshold effect remains a tentative proposition. Additional use frequency data suggest a compound experience difference Task fitP005’s specialization was enterprise UI. LP design may be outside specialization Other participants’ primary specializations also were not LP (P001: SaaS, P002: Healthcare, etc.). Task non-fit is not unique to P005 Largely refutable. However, the distance between enterprise UI and LP may be greater than for others Individual personality This study did not measure personality Not refutable from observational data. Not refutable. Acknowledged as Alternative Explanation Evidence in Present DataRefutabilityJudgment traits/motivationtraitsHowever, P005’s verbalization pattern (external attribution style) is consistent with the conceptual alignment failure explanation unmeasured confound Conceptual alignment failure (this study’s explanation) P005 recorded multiple verbalizations showing mismatch between AI’s and own concept of “LP.” Operational modifications consistently failed Conceptual misalignment confirmed in unit sequence immediately before disengagement, not observed in other successful participants Most consistent explanation within the observed sequence. However, the above confounds cannot be excluded P005 disengagement unit sequence (U3→U4→U5): U3: P005 prompted “Make it a simpler landing page” (L2). AI output: a single-column, spare page. P005: “That’s not what I mean by simple. Our company’s LPs have much more content” → C-semantic. Here the conceptual mismatch regarding “LP” (P005 assumed a content-rich page; AI assumed a minimal page) became manifest. U4: P005 prompted “Add more content, include service description and pricing table” (L2→L3). AI output: multi-element but scattered layout. P005: “Even after explaining this much, it still doesn’t work” → R-surface (superficial dissatisfaction with output). No progression to R-deep. U5 (disengagement unit): P005 prompted “Redo everything. From the beginning” (non- specific instruction). AI output: similar to initial output. P005: “I feel like giving up. I don’t even feel like scoring” → disengagement tendency. Interpretation: The U3-U5 sequence is consistent with a reading in which conceptual misalignment (the meaning gap of “LP”) was never resolved while operational details were continuously added, resulting in frustration accumulation and disengagement. Whether this pattern is associated with conceptual alignment failure or with experience differences and AI use experience threshold effects cannot be discriminated from the present data alone. 6.6.2 General Limitations Sample size: N = 5 was designed from the Information Power perspective, but the “disengagement type” (P005) is represented by only one participant. The breadth and variability of this subcategory could not be established from the present data. Conclusions regarding conceptual alignment are positioned as tentative propositions. Tool dependency: Only Figma Make (version 2025.11) was used. Since the underlying model is not disclosed, replicability with different tools (e.g., Replit AI, Cursor) has not been verified (see Table 4). The impact of output variation was examined in Section 4.3.1, but systematic replicability verification has not been achieved. This study defines replicability not at the level of identical output reproduction but at the level of procedural and analytic replicability— whether lexical oscillation and the ECRT schema can be re-observed in similar generative AI environments using the same protocol and coding approach (see Section 4.3). The opacity of the underlying model was managed as a design constraint, and we judged that this constraint does not directly compromise the validity of conclusions on the grounds that the object of analysis is designers’ interpretive and interactional responses rather than AI output quality. However, the possibility that different model architectures or parameter settings may indirectly influence designers’ cognitive processes cannot be excluded and should be verified in future replications. Cultural and linguistic context: All participants prompted in Japanese (see Section 6.5). The association between lexical oscillation and the high-context nature of Japanese is merely a tendency within the present data scope and is not a causal claim. Task dependency: Limited to the hero section of a B2B SaaS landing page. Lexical oscillation emergence for other design tasks (e.g., dashboard design, mobile app UI) has not been verified. Hawthorne effect: The think-aloud method itself may have increased metacognitive behavior beyond what would occur in natural design practice. Limitations of confidence-satisfaction scores: Scores were used as structured elicitation devices (see Section 3.3) and not for quantitative evaluation. Further assessment of the scoring procedure and its interpretive utility requires additional verification with larger samples. Alternative explanations for lexical oscillation and the ECRT schema observations: Alternative explanations exist for the observed patterns discussed in this study—lexical oscillation, trajectories traced with the ECRT schema, and the qualitative relation between semantic mismatch and deeper reflection—beyond AI’s probabilistic nature. First, the influence of task difficulty. The hero section design requires simultaneous integration of multiple dimensions, and this complexity itself may generate trial-and-error patterns resembling lexical oscillation. However, lexical oscillation was interpreted through trajectory- level analysis rather than fixed necessary-and-sufficient criteria, attending to returns from operational specificity toward ambiguity in context (Table 1), which distinguishes it from mere trial-and-error. Second, the influence of time pressure. The 30-minute session constraint may have prompted “efficient exploration” in participants, potentially elevating lexical oscillation frequency above natural levels. Third, the influence of UI guidance. The UI design of Figma Make itself may have guided specific prompting patterns, raising the possibility that observed lexical oscillation is a behavioral pattern specific to this tool. Fourth, the influence of individual differences. Differences in years of experience, AI use history, and specialization among the five participants (see Table 2) may confound lexical oscillation and the ECRT schema emergence patterns. We did not establish control conditions for causal discrimination against the above alternative explanations, and observed patterns are reported as descriptive trends. Causal verification of lexical oscillation and the ECRT schema requires future confirmatory research including control of tool conditions, manipulation of task difficulty, and stratification by experience level. Member checking: Member checking was not conducted in this study. Analyzed ECRT trajectories and thematic interpretations were not returned to participants for confirmation. This decision was based on two considerations: first, the primary data consisted of real-time think-aloud verbalizations capturing in-the-moment cognitive processes, and retrospective validation risks post-hoc rationalization that could compromise data authenticity (Ericsson & Simon, 1993); second, the abductive coding was guided by sensitizing-concept descriptions (Table 1) rather than participants’ subjective interpretations. Compensating measures included critical-friend dialogue, a reflexive journal, and proactive negative-case analysis (P005). Future replications may incorporate member checking at the thematic level while recognizing its limitations for protocol analysis data. Confirmation bias: The author is an insider researcher with 25 years of UI/UX design practice experience (see positionality in Section 4.1) and faces the risk of overestimating the utility of the ECRT schema. We addressed this reflexively rather than by seeking to eliminate it. First, a critical friend (a doctoral researcher in HCI) reviewed selected coded extracts and challenged candidate themes, helping to surface alternative readings. Second, a reflexive journal documented how the author’s assumptions shaped coding and theme development. Third, the negative case (P005), inconsistent with the developing account, was proactively explored and reported. We do not claim that these measures eliminated interpretive influence stemming from the insider perspective; independent re-analysis by researchers from outside the design domain is encouraged in future replication studies. Data auditability: Raw transcripts are not publicly available due to confidentiality obligations, but the sensitizing-concept descriptions (Table 1), representative unit excerpts, and the audit- trail materials (Supplementary Material) are reported so that third parties can trace the analysis. Additional anonymized data may be requested from the corresponding author (see Section 6.7 for data availability). 6.7 Data Availability Statement Shareable materials (audit-trail artifacts): (a) sensitizing-concept descriptions (Table 1, including boundary notes); (b) the unit segmentation log (the “1 prompt–1 output pair = 1 unit” procedure and the consolidation/exclusion decisions, per Section 4.5); (c) selected coded extracts with analytic memos—representative examples, counterexamples, and disengagement sequences; (d) the analytic memo trail and theme-development records; and (e) critical-friend discussion notes. Consistent with reflexive thematic analysis, these materials are provided to make the analysis auditable, not as reliability evidence. Together they allow third parties to trace the analytic process and to design replications in similar generative AI environments. Non-shareable materials: Raw audio data were deleted in accordance with the deletion policy upon completion of transcription. Screen recording data are stored in encrypted form following anonymization processing. Alternative means: Items (a)–(e) above may be made available as restricted viewing with a non-disclosure agreement (NDA) upon request from the editorial office or reviewers. Additional anonymized data may be requested from the corresponding author within the scope of participant consent. 7. Conclusion 7.1 Summary Through think-aloud sessions with five expert designers (M = 15.4 years of experience), this study found that interaction with generative AI did not proceed as a unidirectional move from vague intent to concrete specification. Instead, design intent was formed and reconfigured through interaction, with designers oscillating across levels of lexical granularity and returning to ambiguity as a resource, within situations where a basic conceptual footing was in place. 7.2 Contributions This study makes three contributions. First, it empirically describes lexical oscillation in expert designers’ interaction with generative AI, including returns from operational specificity to more ambiguous language, and identifies a non-oscillating trajectory (P005’s disengagement) as a negative case that suggests conceptual misalignment as a tentative boundary for future examination. Second, it conceptually reframes prompt-mediated design as a process of design intent formation and treats AI as a non-neutral generative interlocutor. Third, it positions L1– L3 lexical granularity as a sensitizing analytic lens within reflexive thematic analysis and derives three design implications (Conceptual Alignment Verification UI, Consultation Shift Nudge, Lexical Granularity Feedback) for supporting intent formation. 7.3 Future Directions Plans for the second phase include: (1) longitudinal tracking of lexical oscillation skill change; (2) expert-novice comparison; (3) additional analysis of disengagement factors; (4) experimental evaluation of design patterns; (5) cross-linguistic comparison (lexical oscillation in English prompting environments). If design intent is not retrieved ready-made but formed in interaction, then a generative AI— precisely because its outputs are not neutral—can become an occasion for the kind of reflection through which intent takes shape. Disclosure Statement The author reports there are no competing interests to declare. AI Disclosure In accordance with Taylor & Francis policy, the author discloses the use of a generative AI assistant (Claude Opus 4.6, Anthropic) during the preparation and revision of this manuscript. The assistant was used for language editing, structural revision support, wording alternatives, and the preparation of draft response text. All substantive decisions, analysis, interpretation, claims, and references were verified and finalized by the author. The AI assistant was not used for data collection or for the coding and interpretation of the empirical data; all data analysis, thematic interpretation, and the verification of all claims and references were carried out and confirmed by the author, who takes full responsibility for the content and integrity of the manuscript. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. References American Psychological Association. (2017). Ethical principles of psychologists and code of conduct (2002, amended effective June 1, 2010, and January 1, 2017). https://w.apa.org/ethics/code Barad, K. (2007). Meeting the universe halfway: Quantum physics and the entanglement of matter and meaning. Duke University Press. https://doi.org/10.1215/9780822388128 Blumer, H. (1954). What is wrong with social theory? American Sociological Review, 19(1), 3–10. https://doi.org/10.2307/2088165 Boud, D., Keogh, R., & Walker, D. (1985). Reflection: Turning experience into learning. Kogan Page. Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101. https://doi.org/10.1191/1478088706qp063oa Braun, V., & Clarke, V. (2019). Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health, 11(4), 589–597. https://doi.org/10.1080/2159676X.2019.1628806 Buxton, B. (2007). Sketching user experiences: Getting the design right and the right design. Morgan Kaufmann. Creswell, J. W., & Poth, C. N. (2018). Qualitative inquiry and research design (4th ed.). Sage. Cross, N. (2004). Expertise in design: An overview. Design Studies, 25(5), 427–441. https://doi.org/10.1016/j.destud.2004.06.001 Dorst, K., & Cross, N. (2001). Creativity in the design process: Co-evolution of problem–solution. Design Studies, 22(5), 425–437. https://doi.org/10.1016/S0142-694X(01)00009-6 Ericsson, K. A., & Simon, H. A. (1993). Protocol analysis: Verbal reports as data (Rev. ed.). MIT Press. Gaver, W. W., Beaver, J., & Benford, S. (2003). Ambiguity as a resource for design. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’03) (p. 233–240). ACM. https://doi.org/10.1145/642611.642653 Goldschmidt, G. (1991). The dialectics of sketching. Creativity Research Journal, 4(2), 123–143. https://doi.org/10.1080/10400419109534381 Hall, E. T. (1976). Beyond culture. Anchor Books. Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux. Karpathy, A. [@karpathy]. (2025, February 2). There’s a new kind of coding I call “vibe coding” where you fully give in to the vibes [Post]. X. https://x.com/karpathy/status/1886192184808149383 Knoth, N., Tolzin, A., Janson, A., & Leimeister, J. M. (2024). AI literacy and its implications for prompt engineering strategies. Computers and Education: Artificial Intelligence, 6, 100225. https://doi.org/10.1016/j.caeai.2024.100225 Koch, J., Lucero, A., Hegemann, L., & Oulasvirta, A. (2019). May AI? Design ideation with cooperative contextual bandits. In Proceedings of CHI ’19. ACM. https://doi.org/10.1145/3290605.3300863 Kolb, D. A. (1984). Experiential learning: Experience as the source of learning and development. Prentice Hall. Lee, H. P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of CHI ’25. ACM. https://doi.org/10.1145/3706598.3713778 Lieberman, H., Paternò, F., Klann, M., & Wulf, V. (2006). End-user development: An emerging paradigm. In End user development (p. 1–8). Springer. Lincoln, Y. S., & Guba, E. G. (1985). Naturalistic inquiry. Sage. Malterud, K., Siersma, V. D., & Guassora, A. D. (2016). Sample size in qualitative interview studies: Guided by information power. Qualitative Health Research, 26(13), 1753–1760. https://doi.org/10.1177/1049732315617444 Meske, C., Hermanns, T., von der Weiden, E., Loser, K.-U., & Berger, T. (2025). Vibe coding as a reconfiguration of intent mediation in software development: Definition, implications, and research agenda. IEEE Access, 13, 213242–213259. https://doi.org/10.1109/ACCESS.2025.3645466 Nardi, B. A. (1993). A small matter of programming: Perspectives on end user computing. MIT Press. Nonaka, I., & Takeuchi, H. (1995). The knowledge-creating company. Oxford University Press. Polanyi, M. (1966). The tacit dimension. University of Chicago Press. Schön, D. A. (1983). The reflective practitioner. Basic Books. Sidra, S., & Mason, C. (2025). Generative AI in human-AI collaboration: Validation of the collaborative AI literacy and collaborative AI metacognition scales for effective use. International Journal of Human–Computer Interaction, 41, 5084–5108. https://doi.org/10.1080/10447318.2025.2543997 Simkute, A., Tankelevitch, L., Kewenig, V., Scott, A. E., Sellen, A., & Rintel, S. (2025). Ironies of generative AI: Understanding and mitigating productivity loss in human-AI interaction. International Journal of Human–Computer Interaction, 41(5), 2898–2919. https://doi.org/10.1080/10447318.2024.2405782 Suchman, L. A. (1987). Plans and situated actions: The problem of human–machine communication. Cambridge University Press. Tankelevitch, L., Kewenig, V., Simkute, A., Scott, A. E., Sarkar, A., Sellen, A., & Rintel, S. (2024). The metacognitive demands and opportunities of generative AI. In Proceedings of CHI ’24. ACM. https://doi.org/10.1145/3613904.3642902 Tavory, I., & Timmermans, S. (2014). Abductive analysis: Theorizing qualitative research. University of Chicago Press. Vallgårda, A., & Fernaeus, Y. (2015). Interaction design as a bricolage practice. In Proceedings of the Ninth International Conference on Tangible, Embedded, and Embodied Interaction (TEI ’15) (p. 173–180). ACM. https://doi.org/10.1145/2677199.2680594 World Medical Association. (2013). World Medical Association Declaration of Helsinki: Ethical principles for medical research involving human subjects. JAMA, 310(20), 2191–2194. https://doi.org/10.1001/jama.2013.281053 Yang, Q., Steinfeld, A., Rosé, C., & Zimmerman, J. (2020). Re-examining whether, why, and how human-AI interaction is uniquely difficult to design. In Proceedings of CHI ’20. ACM. https://doi.org/10.1145/3313831.3376301 Zamfirescu-Pereira, J. D., Wong, R. Y., Hartmann, B., & Yang, Q. (2023). Why Johnny can’t prompt. In Proceedings of CHI ’23. ACM. https://doi.org/10.1145/3544548.3581388 Supplementary Material An audit trail supporting the reflexive thematic analysis (sensitizing-concept documentation, participant-level occurrence summaries, an exclusive re-categorization, and the original code- by-code agreement computation) is provided as a separate Supplementary Material file. These materials are offered for transparency and auditability only; consistent with reflexive thematic analysis, they are not treated as findings and are not presented as reliability evidence.

Instant research discovery