A Human–AI Collaborative Design Methodology for an Adolescent Smartphone Addiction Management Device: Integrating LLM-AHP, SD-LLM, and Manual Refinement

Abstract

Mitigating smartphone addiction among adolescents remains a significant challenge. Existing anti-addiction products exhibit critical limitations, including vulnerability to circumvention, over-reliance on user self-discipline, and insufficient coercive control. To address these issues, this study proposes a novel product design methodology integrating LLM-AHP (Large Language Model-Analytic Hierarchy Process), SD-LLM (Stable Diffusion-Large Language Model), and manual expert refinement across four iterative stages. The LLM-AHP analysis identified “coercive management independent of self-discipline” and “minimization of human intervention and family conflicts” as the most critical requirements. The final hardware-based device was developed with several key features, including a non-removable case, non-standard internal screws, a triple-verified contact detection device, an emergency call area, and a roller shutter. The workflow of “LLM-AHP → SD-LLM → Manual Refinement” facilitates precise requirement identification, design evaluation, and scheme generation, offering a reusable design framework for other hardware-based children’s products.

1. Introduction

Adolescents are particularly susceptible to smartphone addiction. A study by Zemni [1] indicated that the prevalence of smartphone addiction among medical students at the University of Monastir was 30.2%. Satish [2] found that over 50% of adolescents were addicted to smartphones. Lisa [3] observed that many adolescents spend more than 5 h daily on their phones. A study conducted by Kamollada [4] at Prince of Songkla University revealed that 29.7% of participants used smartphones for 4 to 6 h per day. Smartphone addiction not only increases the risk of deteriorating eyesight but also leads to sleep or psychological issues. Li [5] conducted eye examinations on 523 children and found a positive correlation between daily smartphone usage and the development of myopia. Eliane [6] discovered that for every additional hour of smartphone usage, average sleep duration decreased by 11 min. Jin [7] demonstrated a significant positive correlation between smartphone usage, anxiety, and procrastination. Yang’s [8] study of 510 children indicated a significant positive correlation between smartphone addiction, depression, and autism.

As smartphones are essential tools for communication and accessing information, it is impractical to ban them completely among adolescents [9]. Smartphone usage can be managed through either non-coercive or coercive approaches. Existing methods of non-coercive smartphone management include self-regulation [10,11,12], physical exercise or aerobic activity [13,14], mindfulness therapy [15,16], healthy social relationships [17,18], psychological intervention [19], and medical treatment [20]. However, these non-coercive methods rely on self-discipline and have often demonstrated limited effectiveness [21]. Some researchers believe that coercive approaches are the most effective, such as prohibiting or confiscating smartphones in classrooms [22,23]. Coercive methods can be divided into software-based and hardware-based approaches. In terms of software-based approaches, Vasiliu [24] argued that the applications designed to restrict internet access were effective. Kim [25] developed an application that could automatically shut down smartphones. Wang [26] proposed an anti-addiction quiz system that restricted smartphone access until the user had correctly answered enough questions. However, the software settings can be modified, and passwords could easily be obtained. If the smartphone is rooted, the software will be circumvented [27,28,29,30]. Due to the shortcomings of software, hardware-based methods are considered to be more reliable. There are commercial devices that lock the phone in a box, which can only be opened at the scheduled time [31]. However, such devices rely on self-discipline; if the user does not actively return the smartphone to the box and lock it at the scheduled time, the smartphone will not be effectively controlled. To overcome the shortcomings of relying on self-discipline, several researchers have undertaken further studies. Cai [32] proposed a method of cutting off the power supply of the smartphone screen via a hardware timing unit. However, this requires modification of the smartphone’s circuitry, which is difficult and may affect its integrity. It may also infringe the rights of the manufacturer. Lin [33] proposed an anti-addiction smartphone case that could lock the screen when the locking button on the case was pressed. However, when using Lin’s method, the smartphone might be woken up by incoming calls or message notifications. In addition, existing products cannot address the issue of family conflicts coming from repeated persuasion. In summary, smartphone anti-addiction products still require further improvement and research.

Although some English abstracts of patents relating to smartphone anti-addiction devices are found, there are no published papers on the design methods for these devices. Therefore, this paper focuses on the requirements analysis and design methods for smartphone control products. In recent years, LLMs (Large Language Models) have been applied to the requirements analysis in product design [34]. Jonathan [35] developed a chat-based LLM system that could convert conversations into a list of requirements. Pan [36] used LLM to process multi-source data and established a mapping between design requirements and product features. However, few studies in the existing literature have integrated LLM with AHP in the requirements analysis of product design [37,38,39]. Haeun [40] combined LLM with AHP, using LLM to facilitate AHP modeling and evaluation, and applied this approach to improve performance management in agricultural enterprises. Abdulkadir [41] employed LLM as a virtual expert to evaluate AHP weights and applied it to the site selection of solar power plants. These studies that combine LLM with AHP in other fields provide new perspectives for the design of smartphone anti-addiction devices. There are some papers on using LLM to optimize prompts before the SD (LLM-SD) generation. But no study has utilized LLM to provide suggestions and key design features after the SD initial generation (SD-LLM) [42,43,44]. Due to the limitations of SD and LLM in handling complex and detailed tasks, manual refinement of product design remains essential.

To overcome the shortcomings of existing anti-addiction devices, this study employs an LLM-AHP method to quantify user requirements and utilizes an SD-LLM workflow to generate initial design schemes. Human experts are then invited to refine and improve the optimal scheme. Through this methodology, a novel smartphone anti-addiction device based on a roller shutter mechanism is successfully developed.

2. The Design Method and Process for Anti-Addiction Products

2.1. Research Framework and Process

(1) Requirements of analysis and data preparation based on LLMs: Collected data from interviews, literature, websites, and patents. Used LLMs to generate high-frequency requirement keywords. Designed a questionnaire based on a 5-point Likert scale and conducted a survey.

(2) LLM-AHP weight calculation: Used the LLM-AHP method to classify, calculate weights, and establish a priority ranking based on the requirements obtained from the questionnaires.

(3) SD-LLM generation: Firstly, used SD to generate initial schemes based on the requirement weights; secondly, human experts identified issues in the initial generation; thirdly, used LLM to provide suggestions and design features; finally, used SD to generate new schemes.

(4) Refined the optimal scheme via human experts.

The overall framework is illustrated in Figure 1.

2.2. Requirements Analysis and Data Preparation Based on LLMs

2.2.1. Data Collection

Interviews at a Primary School Entrance

A student majoring in Design conducted interviews with 36 parents at a primary school entrance in Beijing. The interviews focused on the following question: “What are the drawbacks or challenges associated with methods for controlling adolescents’ smartphone usage time?”

Searching Literature Using X-MOL

X-MOL was utilized to retrieve the latest English literature. The URL is: https://www.x-mol.net/advancedSearch (accessed on 2 November 2025). Impact Factor (IF) > 1.0. Time Span, 2019–2026. Keywords: “mobile phone and addiction”, “mobile phone and control”, “mobile phone and reduce usage time”, and “adolescents and mobile phone”.

Searching Patent Abstracts Using the Database of European Patent Office

The database of European Patent Office contains patents from different countries regions such as Europe, the United States, China, and Japan, and provides English abstracts for these patents. The URL is: https://worldwide.espacenet.com/ (accessed on 5 November 2025). Search Type: Advanced Search. Field: “Title, abstract or names”. Keywords: “mobile phone and control”, “mobile phone and reduce usage time”, “adolescents and mobile phone”, and “mobile phone addiction”.

Conducting Web Search Using Microsoft Bing

The international version of Microsoft Bing was utilized. The URL is: https://cn.bing.com/?FORM=BEHPTB (accessed on 6 November 2025). Keyword: “mobile phone control”.

2.2.2. Data Analysis and Requirement Keyword Statistics Based on LLMs

All results obtained in Section 2.2.1 were compiled into a single Word document. ChatGPT-4, DeepSeek-V3.2, Qwen3.5, and Doubao2.0 were employed to analyze this document. The prompt used for the analysis was as follows: “Based on the document, provide user requirements and design features of a smartphone management and control device for adolescents.”

To mitigate the LLMs’ hallucinations and reduce analysis omissions caused by incomplete document reading, each model was queried five times. The repetition rate of the output for each query was required to be less than 80%.

If the LLMs generate $N$ responses for a specific query, and the number of identical responses is $n$, the repetition rate $R$ is calculated using Equation (1):

$$R={\displaystyle \frac{n}{N}}\times 100\%$$

(1)

The analysis results were exported separately, and DeepSeek-V3.2 was used to conduct a statistical analysis of high-frequency requirement keywords.

2.2.3. Questionnaire Design and Survey Implementation

Five students majoring in Design were invited to analyze, review, and organize the design requirements provided by LLMs as described in Section 2.2.2. Subsequently, ten parents of adolescents were invited to provide further feedback on these organized requirements. Based on these results, a questionnaire was designed, and the survey was conducted via Wenjuanxing, a widely used online survey platform in China (https://www.wjx.cn/) (accessed on 22 December 2025).

To quantify the design requirements, a 5-point Likert scale (as shown in

Table 1. Questionnaire format.

Requirement Items	Scale
Significantly reduce smartphone usage time	1	2	3	4	5
Easy to install	1	2	3	4	5
Simple operation	1	2	3	4	5
Coercive control without self-discipline or human Intervention	1	2	3	4	5
……	1	2	3	4	5
Emergency use	1	2	3	4	5
Emergency call	1	2	3	4	5
No need to modify the smartphone	1	2	3	4	5

) was employed. The scale ranges from 1 to 5, where 1 represents “least important” and 5 represents “most important”. Respondents were asked to rate each item based on their own subjective perceptions.

After collecting the questionnaires, the scores and frequencies for each item were calculated. Subsequently, items were selected and retained according to the scores and frequencies.

2.3. LLM-AHP Weight Analysis

2.3.1. AHP Classification Based on the LLM

DeepSeek-V3.2 was employed to analyze, remove redundant items, and classify the questionnaire items from Section 2.2, thereby establishing the objective layer, criteria layer, and method layer. Subsequently, the results were verified through manual review. The LLM prompt was: “Before conducting the AHP analysis, evaluate, remove redundant items, and classify the questionnaire items; define the objective layer, criteria layer, and method layer, and merge similar items.”

2.3.2. AHP Calculations and Analysis

Judgment matrices were constructed for each layer of the AHP. For each requirement factor, a pairwise comparison matrix was established to reflect the relative importance of each factor. During the construction of the judgment matrices, 10 designers conducted pairwise comparisons of the factors and evaluated the relative importance (using a Saaty 1–9 point scale) [45].

The weight vector W for the requirement factors at each level is obtained by normalizing the values in each column, as shown in Equation (2). W_i is the weight of the i-th requirement factor, and a_ij is the relative comparison value between the i-th and j-th factors.

$$W_i={\displaystyle \frac{{\displaystyle \sum _{j=1}^n{a}_{ij}}}{{\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^n{a}_{ij}}}}}$$

(2)

To ensure the consistency of the judgment matrix, a consistency test was conducted. The Consistency Ratio (C_R) was calculated using Equations (3) and (4) to determine whether it was less than 0.1. If C_R < 0.1, the pairwise comparison matrix was considered to have acceptable consistency. λ_max is the maximum eigenvalue of the pairwise comparison matrix, and R_I is the Random Consistency Index.

$$C_I={\displaystyle \frac{\lambda \mathit{max}-n}{n-1}}$$

(3)

$$C_R={\displaystyle \frac{C_I}{R_I}}$$

(4)

Lastly, the items were ranked based on the comprehensive weights.

2.4. Scheme Generation

SD was utilized to generate the initial schemes based on the comprehensive requirement weights derived from Section 2.3. The five students majoring in Design, as described in Section 2.2.3, evaluated the initial generation according to each requirement and identified existing problems based on their knowledge. DeepSeek-V3.2 was then employed to provide suggestions and key design features for these issues. The prompt was: “Provide suggestions and key design features for the problems identified in the initial generation according to the attachment.”

Based on the suggestions and key design features provided by DeepSeek-V3.2, the prompts were revised, and SD was used again to generate four new design schemes.

Subsequently, the five students majoring in Design selected the optimal design scheme from the four new schemes. The methodology for selecting the optimal scheme consisted of three stages: evaluating the advantages and disadvantages of each scheme, reaching a consensus via team discussion, and conducting a final vote, where the scheme with the most votes was selected as the optimal one.

2.5. Refinement of the Optimal Scheme

Since LLMs and SD are inherently limited in addressing complex problems with ambiguous boundaries and providing exhaustive details, human experts were required to refine the materials, parameters, structures, circuits, and other technical specifications based on the optimal scheme to ensure its practical feasibility.

3. Design Practice and Results

3.1. Results of Requirements Analysis and Data Preparation Based on LLMs

Through interviews with 36 parents, a total of 50 feedback items were collected. Additionally, 82 highly relevant academic papers, 31 patents, and 221 web pages were gathered to supplement the dataset. The five students majoring in Design assisted in organizing the LLM-generated data, and ten parents were invited to provide feedback to refine the questionnaire content. Ultimately, 423 valid questionnaires were collected, ensuring that the data sources were sufficiently representative. The questionnaire demonstrated excellent internal consistency (Cronbach’s α = 0.951), confirming its suitability for subsequent statistical analysis.

Based on the high-frequency keyword analysis in Section 2.2.2, a word cloud of user requirements was generated (Figure 2). The four most critical requirement items identified were: “prevent addiction”, “coercive control”, “reduce time”, and “reduce conflicts.” The subsequent eight important items included: “independent of self-discipline”, “cannot be circumvented”, “emergency call”, “reduce human intervention”, “timer function”, “no proof”, “reduce persuasion”, and “emergency use”.

The survey questionnaire comprised a total of 37 requirements. The frequency distribution of the scores for these requirements is presented in Figure 3. The frequency distribution reflects the degree of consensus among different respondents regarding each item. Notably, from item 30 onwards, the curves for the 3-point, 4-point, and 5-point ratings overlapped significantly, indicating a substantial divergence in respondents’ scores for items 30 to 37.

The cumulative percentage distribution of the rating frequencies is illustrated in Figure 4. One-point and two-point ratings represent “unimportant”, while three-point indicates “uncertain”. From item 30 onwards, the cumulative percentages of 1-point, 2-point, and 3-point ratings exceeded 40%, whereas those of 4-point and 5-point ratings fell below 60% (the acceptance threshold). Given that 4-point and 5-point ratings indicate importance and strong support, items 30–37 failed to receive sufficient endorsement from the majority of respondents. Consequently, this study excluded items 30–37 and retained the top 29 items for subsequent analysis.

Figure 5 presents the average scores and percentage of 5-point ratings, displaying only the top 29 items. The average score reflects the level of support from respondents. For these top 29 requirements, the average scores ranged from 3.66 to 4.04. All average scores exceeded 70% of the maximum score of 5 points. An absolute majority of the respondents strongly supported the top 29 requirements. Furthermore, compared to the LLM-based statistical word cloud in Figure 2, Figure 5 provides more specific descriptions and supplementary details based on the word cloud’s content.

3.2. Results of LLM-AHP Weight Calculation

As shown in Figure 6, DeepSeek-V3.2 has categorized the questionnaire items into three layers: Objective Layer (A), Criteria Layer (B), and Method Layer (C). Criteria Layer (B) included: “Coercive management and control capabilities (B1)”, “Flexible emergency and response capabilities (B2)”, “User-friendly and convenient (B3)”, and “Appearance and structure (B4)”. Additionally, Method Layer (C) included 24 items.

Table 2. Results of the B1–B4 weight calculations.

Items	B1	B2	B3	B4	Feature Vectors	Weight
B1	1	5.7	6.9	7.7	4.2132	0.6512
B2	0.1754	1	3	4.9	1.2541	0.1941
B3	0.1449	0.3333	1	2.9	0.6522	0.1009
B4	0.1299	0.2041	0.3448	1	0.3511	0.0538

,

Table 3. Results of the C1–C10 weight calculations.

Items	C1	C2	C3	C4	C5	C6	C7	C8	C9	C10	Feature Vectors	Weight
C1	1	3.1	3.7	5.2	5.2	8.7	9	8.9	8.7	9	5.211	0.3421
C2	0.323	1	3.1	3.7	4.7	7.7	7.7	7.8	7.8	8.1	3.482	0.2286
C3	0.27	0.323	1	2.9	4.1	6.2	6.5	6.3	6.2	6.7	2.265	0.1487
C4	0.192	0.27	0.345	1	3.6	5	5.1	5	5.1	5.8	1.642	0.1078
C5	0.192	0.213	0.244	0.278	1	3.4	3.9	3.9	3.9	4.2	0.987	0.0648
C6	0.115	0.13	0.161	0.2	0.294	1	1.8	2.2	2.3	2.7	0.452	0.0297
C7	0.111	0.13	0.154	0.196	0.256	0.556	1	1.6	2	2.4	0.368	0.0242
C8	0.112	0.128	0.159	0.2	0.256	0.455	0.625	1	1.6	2.2	0.335	0.022
C9	0.115	0.128	0.161	0.196	0.256	0.435	0.5	0.625	1	1.7	0.276	0.0181
C10	0.111	0.123	0.149	0.172	0.238	0.37	0.417	0.455	0.588	1	0.214	0.0140

,

Table 4. Results of the C11–C12 weight calculations.

Items	C11	C12	Feature Vectors	Weight
C11	1	1.1	1.0488	0.5119
C12	0.909	1	0.9535	0.4881

,

Table 5. Results of the C13–C21 weight calculations.

Items	C13	C14	C15	C16	C17	C18	C19	C20	C21	Feature Vectors	Weight
C13	1	4	6	8.2	8.8	9	9	9	9	5.682	0.4263
C14	0.25	1	3.2	5.4	6.6	7.6	7.8	7.6	7.8	3.215	0.2412
C15	0.167	0.313	1	3.2	4.4	5.2	5.2	5.2	5.6	1.742	0.1307
C16	0.122	0.185	0.313	1	3.2	4.2	4.4	4.4	4.6	0.985	0.0739
C17	0.114	0.152	0.227	0.313	1	3.2	3.4	3.4	3.8	0.614	0.0461
C18	0.111	0.132	0.192	0.238	0.313	1	1.8	2.4	2.6	0.358	0.0269
C19	0.111	0.128	0.192	0.227	0.294	0.556	1	1.6	2.2	0.285	0.0214
C20	0.111	0.132	0.192	0.227	0.294	0.417	0.625	1	1.8	0.246	0.0185
C21	0.111	0.128	0.179	0.217	0.263	0.385	0.455	0.556	1	0.201	0.0151

and

Table 6. Results of the C22–C24 weight calculations.

Items	C22	C23	C24	Feature Vectors	Weight
C22	1	2	7.2	2.433	0.6021
C23	0.5	1	3	1.145	0.2832
C24	0.139	0.333	1	0.461	0.1147

present the judgment matrices and local weights. In Table 2, “Coercive management and control capabilities (B1)” (65.12%) accounted for the largest proportion, followed by “Flexible emergency and response capabilities (B2)” (19.41%), then “User-friendly and convenient (B3)” (10.09%), and finally “Appearance and structure (B4)” (5.38%).

The C_R values for Layer A and Layers B1–B4 were 0.0304, 0.0419, 0.0000, 0.0438, and 0.0181, respectively. All C_R values were less than 0.1, and all satisfied the consistency requirement.

Table 7. Comprehensive weight calculations.

Items	Requirements	Comprehensive Weights	Ranking
C1	Coercive management and control for smartphone independent of self-discipline (C1)	0.2228	1
C2	Minimize human intervention, verbal persuasion, and family conflicts (C2)	0.1489	2
C11	Emergency use (e.g., scanning a QR (Quick Response) code to pay for purchases) and coercive covering screen immediately after use (C11)	0.0994	3
C3	Cannot be circumvented or deceived (C3)	0.0968	4
C12	Emergency call (C12)	0.0947	5
C4	Coercive covering screen at the scheduled time (C4)	0.0702	6
C13	Sturdy and durable (C13)	0.043	7
C5	Remote control (C5)	0.0422	8
C22	Beautiful and attractive design (C22)	0.0324	9
C14	High stability (C14)	0.0243	10
C6	Timer function (C6)	0.0193	11
C7	Automatically shut down when the battery is low (C7)	0.0157	12
C23	Comfortable grip with an ergonomic design (C23)	0.0152	13
C8	A device independent of the smartphone (C8)	0.0143	14
C15	Low power consumption (C15)	0.0132	15
C9	Difficult to damage; attempts to damage will trigger an alert (C9)	0.0118	16
C10	Warning for improper use (C10)	0.0091	17
C16	Lightweight, compact, and portable (C16)	0.0075	18
C24	Serve as a case to protect the phone (C24)	0.0062	19
C17	Simple to operate and easy to use (C17)	0.0046	20
C18	Easy to install (C18)	0.0027	21
C19	No need to modify the smartphone (C19)	0.0022	22
C20	Compatible with various phones, can be customized based on phone shape (C20)	0.0019	23
C21	The charging port is compatible with the phone (C21)	0.0015	24

shows the comprehensive weight ranking, while Figure 7 illustrates the trend of multiples relative to the minimum weight. According to Figure 7, the weights are generally divided into five intervals: 1–2 (very strong weight), 3–6 (strong weight), 7–11 (moderate weight), 12–16 (weak weight), and 17–24 (very weak weight). The top two requirements were: “Coercive management and control for smartphone independent of self-discipline” (22.28%), and “Minimize human intervention, verbal persuasion, and family conflicts” (14.89%). The next four items were: “Emergency use and coercive covering screen immediately after use” (9.94%), “Cannot be circumvented or deceived” (9.68%), “Emergency call” (9.47%), and “Coercive covering screen at the scheduled time” (7.02%). It could be seen that users place the highest priority on coercive control, minimizing human intervention, reducing conflicts, anti-circumvention, emergency use, and coercive covering screen. From the 7th rank, the value decreased to less than 5%, and from the 17th rank, it dropped to under 1%.

3.3. Results of the Scheme Generation

In SD prompts, the original weights must be converted into relative weights [46]. Relative weight values exceeding 2.0 may lead to image distortion. An optimal range for relative weights is generally considered to be between 0.5 and 1.7. However, as illustrated in Figure 7, the maximum original weight (0.2228) was 149 times greater than the minimum (0.0015). Consequently, it was impossible to normalize all 24 requirements within the target range. Therefore, this study adopted a “stratified weight” strategy based on the five intervals shown in Figure 7. For each layer of parentheses “(tag)”, the relative weight is multiplied by 1.1; for each layer of curly braces “{tag}”, the weight is multiplied by 1.05; and for each layer of square brackets “[tag]”, the weight is multiplied by 0.9.

Table 8. Prompts used for the initial generation of SD.

Positive Prompts Used for the Initial Generation of SD

Negative Prompts Used for the Initial Generation of SD

(((Smartphone anti-addiction device))), (((Coercive management and control for smartphone independent of self-discipline))), ((Minimize human intervention, verbal persuasion, and family conflicts)), {{{Emergency use, scanning a QR code to pay for purchases, coercive covering screen immediately after use}}}, {{Cannot be circumvented or deceived}}, (Emergency call), {Coercive covering screen at the scheduled time}, Sturdy and durable, Remote control, Beautiful and attractive design, High stability, Timer function, [Automatically shut down when the battery is low], [Comfortable grip with an ergonomic design], [[A device independent of the smartphone]], [[Low power consumption]], [[Difficult to damage, attempts to damage will trigger an alert]], [[[Warning for improper use]]], [[[Lightweight, compact, and portable]]], [[[Serve as a case to protect the phone]]], [[[Simple to operate and easy to use]]], [[[Easy to install]]], [[[No need to modify the smartphone]]], [[[Compatible with various phones, can be customized based on phone shape]]], [[[The charging port is compatible with the phone]]]

Ugly, deformed, blurry, low resolution, messy lines, extra parts, broken screen, text, watermark, cartoon, anime, illustration, unrealistic, plastic feel, too bright, too dark, disorganized, open box, password lock, ordinary phone case, ordinary phone management and control

shows the prompts used for the initial generation of SD.

In

Table 9. Combined use of SD-LLM.

Items	Issues Identified By The Design Students In The Initial Generation of SD	Use LLM to Provide Suggestions and Design Features for The Initial Generation of SD
C1	The phone is stored in a box and must be returned to it; if it is not returned automatically, the device cannot control the phone	Coercive covering screen with a roller shutter at the scheduled time
C2	Lack details	The device is integrated with the phone and cannot be removed; the action cannot be undone
C11	Cannot be opened until the scheduled time	In an emergency the device can be opened for two minutes, it can only be used for emergency purposes five times per 24 h period
C3	The device uses a password system; the password may be stolen or obtained	Using two non-standard internal triangular screws to fix the phone case
C12	The box must be opened to make a call	Reserve a section at the bottom of the device that does not cover the screen, allowing for phone calls
C4	The device uses a door-opening mechanism; the door takes up space when opening and closing	Covering screen with a roller shutter
C13	The design uses plastic and is fully enclosed	Aluminum alloy construction with a hollowed frame design for easy access to ports and components
C5	No communication module was designed	Equipped with an independent 5G module
C22	The color and shape are unattractive.	Beautiful color and attractive design
C14	Lack details	Extremely reliable roller shutter mechanisms and control systems
C6	Uses a timer-based switch	Set the time using the software instead of the switch
C7	Lack details	The device will automatically shut down to prevent circumventing when the device is low on power
C23	Lack details	Referencing the ergonomics of mobile phones
C8	The device is separate from the phone	The device is fixed to the outside of the phone as a whole and is non-removable
C15	Includes a display screen, which consumes power	The device has no display screen
C9	Lack details	The device can only be disassembled from the two screw points
C10	Designs an external alarm light	If anyone attempts to prevent the roller shutter from closing, it will sound an alarm and vibrate, distracting those watching the screen
C16	The device is very large.	Miniaturization of components such as batteries and motors
C24	Simply designed as a traditional phone case.	Serve as an aluminum alloy case to protect the phone
C17	Lack details	The device has no buttons
C18	Lack details	Tightening two screws is the only step in the installation process
C19	Lack details	No need to modify the buttons or circuit
C20	Lack details	Compatible with various phones, can be customized based on phone shape
C21	Lack details	The charging port is compatible with the phone

, the second column outlines the issues identified by the five students majoring in Design in the initial generation of SD, the third column shows the suggestions and design features for the initial generation of SD based on DeepSeek-V3.2.

The issues in the initial generation of SD were as follows: the smartphone must be placed back in the case, passwords are easily stolen, it takes up a lot of space, it is not convenient for emergencies, etc. DeepSeek-V3.2 provides suggestions and key design features: a non-removable case, non-standard triangular internal screws to prevent circumvention, an emergency call area at the bottom, a roller shutter to cover the screen, etc.

Table 10. Prompts used for the second generation of SD.

Positive Prompts Used for the Second Generation of SD

Negative Prompts Used for the Second Generation of SD

(((Smartphone anti-addiction device))), (((Coercive covering screen with a roller shutter at the scheduled time))), ((The device is integrated with the phone and cannot be removed; the action cannot be undone)), {{{In an emergency the device can be opened for two minutes, it can only be used for emergency purposes five times per 24-h period}}}, {{Using two non-standard internal triangular screws to fix the phone case}}, (Reserve a section at the bottom of the device that does not cover the screen, allowing for phone calls), {Covering screen with a roller shutter}, Aluminum alloy construction with a hollowed-out frame design for easy access to ports and components, Equipped with an independent 5G module, Beautiful color and attractive design, Extremely reliable roller shutter mechanisms and control systems, Set the time using the software instead of the switch, [The device will automatically shut down to prevent circumventing when the device is low on power], [Referencing the ergonomics of mobile phones], [[The device is fixed to the outside of the phone as a whole and is non-removable]], [[The device has no display screen]], [[The device can only be disassembled from the two screw points]], [[[If anyone attempts to prevent the roller shutter from closing, it will sound an alarm and vibrate, distracting those watching the screen]]], [[[Miniaturization of components such as batteries and motors]]], [[[Serve as a aluminum alloy case to protect the phone]]], [[[The device has no buttons]]], [[[Tightening two screws is the only step in the installation process]]], [[[No need to modify the buttons or circuit]]], [[[Compatible with various phones, can be customized based on phone shape]]], [[[The charging port is compatible with the phone]]]

Ugly, deformed, blurry, low resolution, messy lines, extra parts, broken screen, text, watermark, cartoon, anime, illustration, unrealistic, plastic feel, too bright, too dark, disorganized, open box, password lock, ordinary phone case, ordinary phone management and control

shows the prompts used for the second generation of SD. All suggestions and key design features have been incorporated into the prompts used for the second generation according to their respective weights.

Figure 8a–d show the four design schemes presented in the second generation of SD. All four schemes essentially employed a roller shutter covering method and could achieve the objective of coercive management and control. In schemes 1 and 2, the roller was relatively large, which partly covered the upper part of the screen. In scheme 3, the roller partly covered the middle part of the screen. Scheme 4 placed the roller in the right compartment, ensuring it did not cover the screen. Scheme 4 had a more logical design and served as a basis for further refinement. A vote was conducted among the five students majoring in Design, and Scheme 4 received unanimous support (five votes). Because the prompt did not include the phrase “the roller must not interfere with the phone screen under any circumstances”, the impractical roller positions appeared in Schemes 1–3.

3.4. Detailed Results for the Core Components of the Optimal Scheme

3.4.1. Detailed Results for the Device Case

Figure 9 shows the device case. The case was made of 0.3 mm-thick aluminum–magnesium alloy. It included two compartments: a phone compartment and a roller shutter compartment. The phone compartment was designed to hold and secure the smartphone. Once the phone was secured, it was difficult to remove without damaging the case or using a non-standard wrench. The roller shutter compartment was designed to store the battery, the roller shutter, the motor, the control module, etc. To avoid covering the phone’s camera, buttons, charging port, SIM card slot, earpiece, and microphone, the case had an open-frame design. A 2 mm-wide edge extended horizontally into the phone compartment along each side, and a total of eight edges were used to secure the phone. The device case could be flexibly customized to accommodate different smartphones. The upper part of the phone compartment had a roller shutter slot designed to guide the movement and secure the roller shutter. The slot had a depth of 2 mm and was made of 0.3 mm-thick PVC. The smartphone compartment was equipped with two non-standard internal triangular screws, positioned as shown in Figure 9. After the phone was inserted from the left side, it would be locked in the phone compartment by these two screws. The screw sockets had a non-standard, irregular internal triangular design. Once parents had tightened the screws, they could hide the wrench in a location unknown to the child or discard it permanently. Without a matching tool, the phone could only be removed by damaging the case’s structure. Figure 10 is a material and scene rendering of the device.

3.4.2. The Contact Detection Device of Roller Shutter Head

The contact detection device for the roller shutter head was installed in the roller shutter slot on the left side of the mobile phone compartment, as shown in Figure 9 and Figure 11. The contact detection device incorporated two electrical circuits: Circuit A and Circuit B.

Figure 11a shows that the roller shutter head had not yet reached its position. In Circuit A, components 9 and 12 were connected, whereas in Circuit B, components 13 and 14 were disconnected. When Circuit A was closed, and Circuit B was open, the roller shutter had not yet reached its position.

Figure 11b shows that the roller shutter head had reached its position. In Circuit A, the insertion of the insulated roller shutter caused 9 and 12 to disconnect. In Circuit B, the insertion of the bottom electrical contact plate of the roller shutter head caused 14 and 16 to connect. Circuit A was open, Circuit B was closed, and the roller shutter had reached its position.

To prevent deception or circumvention, the contact detection device employed a triple-verified design.

While the roller shutter was moving, if an insulating object, such as cloth, paper, or a finger, was inserted into Circuit A, preventing the shutter from closing, but Circuit B in Figure 11b was not connected, the device would detect a circumvention attempt and trigger a warning.

While the roller shutter was moving, if a metal object was inserted into Circuit B, Circuit B in Figure 11b would be connected. However, since Circuit A was not disconnected by the insulating roller shutter, the device would still trigger a warning.

While the roller shutter was moving, if an insulating object was used to disconnect Circuit A, a metal object was used to connect Circuit B. However, the extension length and moving time of the roller shutter did not match the preset values, and the device would still trigger a warning.

This triple-verified design effectively prevented attempts to deceive or circumvent the roller shutter.

3.4.3. The Roller Shutter and Drive Unit

The roller shutter and its drive unit are shown in Figure 9, Figure 12 and Figure 13. The width of the roller shutter was three-quarters of a standard smartphone screen, with approximately one-quarter of the bottom reserved for emergency calls. To minimize storage space, the roller shutter was made of a 0.3 mm-thick steel sheet. This steel construction prevented users from removing the shutter from the slot or damaging it. Insulating material was applied to the roller shutter surface to provide electrical insulation for the contact detection device. To facilitate rolling, the roller shutter included large and small segments. The width of the large segments was 2 mm, and the width of the small segments was 1 mm. All large and small segments were connected by flexible steel wires. A hidden electrical contact plate was installed at the bottom of the roller shutter head to prevent deception or circumvention. Figure 13 is a transparent view of the roller shutter and its drive unit, showing various internal modules.

3.4.4. Battery Module

The battery module is shown in Figure 13. When the battery level drops below 15%, a charging alert is triggered. If the battery is not charged in time, the roller shutter would automatically cover the screen. This design prevents users from circumventing by intentionally draining the battery.

3.4.5. Control Module

The control module is shown in Figure 13. The control module includes a central processing unit (CPU), a voice recognition module, an alarm horn, a vibration module, a communication module, a remote control module, a timer, a relay module, and a switching mechanism for the roller shutter.

In an emergency, such as when scanning a QR code to pay for purchases, users would use the voice command “Emergency Use”. Upon receiving this command, the control module would open the roller shutter, which would then automatically close after two minutes. To limit usage time, this function could only be activated five times within a 24 h period.

If the phone was used improperly, the control module would trigger a continuous, loud alarm exceeding 70 decibels. The alarm sounds from multiple locations on the device, making it impossible for users to muffle the sound by blocking a specific area. Simultaneously, the device would vibrate continuously, making it impossible for users to focus on the screen or use the phone comfortably.

4. Discussion

This study employed a framework of “LLM-AHP → SD-LLM → Manual Refinement” to address the core challenges in the hardware design for preventing smartphone addiction. The main contribution of this paper is not the creation of new AI algorithms or technical models, but the integration of a fresh design methodology and workflow [47].

4.1. Comparative Analysis with Previous Studies

Table 11. Comparison of the suggested methodology with current AI-assisted or traditional product design approaches.

No.	Dimensions	Existing Studies Approaches	Proposed Methodology	Unique Features and Benefits
1	Management Strategy and Effectiveness	Software-based solutions: Vulnerable to being rooted, bypassed, or password-cracked.	Physical coercive control based on a roller-shutter mechanism: Combines coercive control with non-destructive hardware design.	Independence from self-discipline: Physically covers the screen, overcoming the vulnerabilities of software and the self-discipline reliance of existing hardware.
				Hardware integrity: Requires no internal circuit modification, avoiding infringement risks.
		Existing hardware devices (e.g., timer boxes, circuit cut-offs): Either still rely on user self-discipline or require modification of the smartphone’s circuitry, posing integrity and infringement risks.		Conflict reduction: Automated operation minimizes family conflicts caused by repeated verbal persuasion.
2	Requirements Analysis	Manual AHP methods: Rely on experts’ subjective stratification and are inefficient in processing massive datasets.	LLM-AHP quantitative analysis framework: Utilizes LLM for semantic understanding, redundancy filtering, and hierarchical structuring, combined with AHP for quantitative calculation.	High objectivity: LLM assistance ensures comprehensive and mutually exclusive categorization with high logical consistency.
				High reliability: Achieves a Cronbach’s α of 0.951.
		Current LLM applications: Mostly used for simple conversation-to-requirement conversion or in other fields (e.g., agriculture, site selection).		Domain expansion: First application of LLM-AHP to hardware product requirement quantification, filling a gap in anti-addiction device design methodologies.
3	Design Workflow	LLM-SD: Uses LLM to optimize prompts before SD generation. SD lacks mechanical logic, often resulting in visually appealing but structurally unfeasible designs.	SD-LLM workflow: Generates initial visual prototypes via SD, then employs LLM as a post-processing refinement agent for structural optimization and detail correction.	Enhanced structural logic: Leverages LLM’s reasoning capabilities to compensate for SD’s shortcomings in mechanical structure design.
		Pure manual design: Low efficiency and long iteration cycles.		Iteration efficiency: Enables a rapid closed-loop transition from “visual concepts” to “engineering details,” accelerating the hardware product development process.

shows the comparison of the suggested methodology with current AI-assisted or traditional product design approaches.

4.1.1. Rationale for Moderate Coercive Management

The literature cited in Section 1 has reported numerous non-coercive approaches; however, their effectiveness remains limited [10,11,12,13,14,15,16,17,18,19,20,21]. Given the behavioral inertia commonly observed in adolescents, moderate coercive management is considered essential. Although some existing studies have supported the implementation of coercive control, those studies lack quantitative data to justify its rationality [22,23,24]. For the first time, this study provided quantitative analytical results regarding coercive control using the LLM-AHP method. As shown in Table 2, Table 3, Table 4, Table 5, Table 6 and Table 7, the quantitative analysis indicates that coercive control capability B1 is the primary requirement (weight: 65.12%). This is specifically characterized by C1 (coercive control independent of self-discipline, weight: 0.2228), C3 (inability to be circumvented or deceived, weight: 0.0968), and C4 (coercive screen covering at scheduled times, weight: 0.0702). These quantitative findings reveal that respondents place coercive control at the highest priority, thereby providing quantitative support for the necessity of such methods discussed in the previous literature.

4.1.2. Methodological Transfer: Extending LLM-AHP to Hardware Product Design

Previous studies primarily relied on manual methods for constructing AHP hierarchies, heavily depending on experts’ subjective stratification and categorization [48]. This study employed LLMs to facilitate hierarchical structuring in AHP, taking advantage of their superior semantic understanding and data processing capabilities. Compared to human experts, LLMs can efficiently process extensive datasets and consolidate similar items, ensuring that requirement categories are both comprehensive and mutually exclusive. By utilizing LLM (e.g., DeepSeek-V3.2) for redundancy filtering and hierarchical classification (Figure 6), this research ensured logical consistency in the judgment matrix (C_R < 0.1) and high reliability (Cronbach’s α = 0.951), thereby significantly enhancing the objectivity of the requirements analysis. While existing studies have applied LLM-AHP in agricultural management and site selection, this research expands its application boundary to industrial product requirement quantification, providing a standardized quantitative pathway for hardware product design [40,41].

4.1.3. Methodological Innovation: Shifting from “LLM-SD” to “SD-LLM” for Generation Iteration

While SD is competent in generating rapid visual prototypes based on weight-guided prompts, it lacks logical reasoning capabilities for mechanical structures. Conversely, LLM excels at structural optimization suggestions but cannot independently visualize design concepts. In this study, LLM was employed to refine generation details following the initial SD output, thereby enhancing the precision of subsequent image generation. This approach serves as a further innovation of previous literature that utilized LLM prior to SD (LLM-SD) [49]. Crucially, by shifting the process from a traditional “LLM-SD” paradigm to a novel “SD-LLM” workflow, this study employed the LLM as a post-processing refinement agent, effectively accelerating the generation iteration process in hardware product design.

4.2. Novel Findings and Implications

4.2.1. Discrepancy Between Questionnaire and Weight Ranking

The final weight ranking, derived from the Saaty 1–9 point scale, differed significantly from the original questionnaire results based on a 5-point Likert scale. As shown in Figure 5, the ratio between the maximum and minimum average scores was merely 1.1, whereas in Figure 7, the corresponding ratio for weights reached 149. This discrepancy can be attributed to the fundamental difference in evaluation mechanisms. The original questionnaire did not involve exclusive scoring or direct comparison, allowing respondents to assign maximum scores to all items indiscriminately. In contrast, the expert scoring process required pairwise comparisons, which forced respondents to differentiate and prioritize among the requirements. This reveals distinct psychological mechanisms across various evaluation dimensions. In product design, comparative methods are essential for distinguishing and prioritizing core requirements over superficial preferences. By translating non-linear, ambiguous, and intertwined user inputs into clear, quantifiable weights through mathematical logic (eigenvectors), AHP provides direct and practical guidance for actual hardware design work. The 5-point Likert scale plays an outstanding role in large-scale requirement collection, ensuring that no potential requirements are overlooked. This study first utilizes the 5-point Likert scale to extensively collect initial requirements. Subsequently, the AHP method is employed to enforce pairwise comparisons. The advantages of these two methods are highly complementary.

4.2.2. The Critical Importance of Minimizing Human Intervention

Superficially, one might assume that “Coercive management (C1)”, “Emergency use (C11)”, “Cannot be circumvented (C3)”, “Emergency call (C12)”, “Coercive covering screen (C4)”, or “Automatically shut down (C7)” represent the most critical factors. However, this study revealed that minimizing human intervention (C2, weight: 0.1489) was the second core requirement.

C2 includes three dimensions: “human intervention”, “verbal persuasion”, and “family conflicts”. Analysis indicates that the factors directly correlated with C2 include: “Coercive management and control for smartphone independent of self-discipline (C1) 0.2228”, “Cannot be circumvented or deceived (C3) 0.0968”, “Coercive covering screen at the scheduled time (C4) 0.0702”, “Timer function (C6) 0.0193”, “Automatically shut down when the battery is low (C7) 0.0157”, “Difficult to damage; attempts to damage will trigger an alert (C9) 0.0118”, “Warning for improper use (C10) 0.0091”, etc.

Excellent product design must reflect users’ psychological dynamics. Key factors in minimizing human intervention include automation, coercive management, hardware-based mechanisms, and scheduled timing. This demonstrates that the AHP model successfully uncovers users’ deepest subconscious emotional requirements, rather than merely cataloging surface-level functionalities. Ultimately, this validates the effectiveness of the LLM-AHP methodology—proving that AI is capable of not only processing objective data but also accurately extracting complex psychological and sociological pain points inherent in human behavior.

4.3. Limitations of the Study

4.3.1. Absence of Long-Term Efficacy Validation

A primary limitation of this study is the lack of long-term efficacy validation, as the prototype has not yet been deployed in real-world ecological settings to assess its sustained control performance. Consequently, longitudinal tracking studies evaluating the device’s impact on vision improvement and sleep quality have not been conducted. Future research should prioritize field testing to verify these long-term health outcomes.

However, to bridge the gap between theoretical design and practical application, the study temporarily conducted a desk-based evaluation focusing on the blueprints, detailed specifications, core functions, and parameters. To comprehensively reflect both the users’ willingness to adopt the product and the manufacturing feasibility, 15 participants (5 parents, 5 students, and 5 product manufacturing workers) were invited to rate the design. All scores reached or exceeded 8.0 (out of 10), indicating high theoretical feasibility and user acceptance.

As shown in

Table 12. Desk-based evaluation of the design.

Acceptance	Parents	Students	Feasibility	Product Manufacturing Workers
Appearance	8.4	7.8	Material	9.8
Structural	8.4	8.4	Structure	8.6
Functional	8	8.4	Function	8.4
Willingness to Use	9.8	8.4	Circuit	9.6

, this desk-based evaluation assessed acceptance, structure, willingness to use, core functions, and product feasibility. The high ratings from both end-users and manufacturing workers provide supplementary evidence for the theoretical validity and practical potential of the proposed design.

4.3.2. Questionnaire Reliability and Sample Scope

Although the questionnaire demonstrated acceptable reliability (Cronbach’s α = 0.951), it was only distributed via the Wenjuanxing platform. Consequently, potential response biases, such as careless responding, could not be entirely ruled out. In psychometric analysis, values above 0.9 may also suggest a high degree of redundancy among the questionnaire items. It is necessary to streamline the questionnaire (e.g., by removing overlapping questions) in future studies. Furthermore, this study only conducted basic reliability testing, frequency inconsistency analysis, and descriptive statistical analysis (Figure 3, Figure 4 and Figure 5). Future research requires additional methods to rigorously verify the scientific validity of the questionnaire. Additionally, the interviews were confined to a single primary school in Beijing (Section 2.2.1); future studies need to expand the sample size across diverse urban and rural regions to enhance generalizability.

4.3.3. Limitations of SD as a Generation Tool

Although newer commercial image generation systems, such as Gemini 3.0, have emerged with superior capabilities, they often lack transparency and local deployability. During the SD generation process in this study, initial weights could not be directly mapped and had to be substituted with relative weights; furthermore, SD lacks the direct mathematical capacity to fully visualize these complex weighted items. Nevertheless, SD remains a highly viable option for academic research due to its open-source nature, local deployment capability, and advanced customizability (e.g., ControlNet, LoRA). These features provide researchers with partial but essential control over detail generation, which is often unattainable with closed-source models.

4.3.4. Limitations of the Construction of the Judgment Matrices

The construction of the judgment matrices and the pairwise comparisons for the AHP calculations were performed by 10 designers. The absence of domain experts such as child psychologists, pediatricians, educators, and parents undermines the validity of the psychological and behavioral weightings.

To address this issue, three child psychologists, two pediatricians, and two educators (primary school teachers) were invited to evaluate the weights and rankings (Table 2, Table 3, Table 4, Table 5, Table 6 and Table 7) derived from the 10 designers’ scores to determine whether there are significant differences. A five-tier evaluation scale (A to E) with fine-tuning (+/-) was adopted: A (Agree), B (Mostly agree), C (Slightly agree), D (Disagree), and E (Strongly disagree).

As shown in

Table 13. Evaluation of Table 2, Table 3, Table 4, Table 5, Table 6 and Table 7 from domain experts.

Items	Table 2	Table 3	Table 4	Table 5	Table 6	Table 7
Child psychologist 1	B+	B-	B-	B-	B-	C+
Child psychologist 2	B	B-	B-	C+	B-	B
Child psychologist 3	B	C+	B	C+	B+	B
Pediatrician 1	B-	B	B+	B	B	B-
Pediatrician 2	C+	B-	B	B+	C+	B-
Educator 1 (Primary school teacher)	B+	C+	B-	B	B-	C+
Educator 2 (Primary school teacher)	B	B+	B	B-	B	B-

, a total of 42 evaluation items (7 experts × 6 tables) were assessed. A total of 34 items (80.95%) were rated at Level B- (Mostly agree) or above, while the remaining 8 items were rated at Level C+ (Slightly agree). This indicates a moderate to high level of agreement between the experts’ criteria and the designers’ weightings, generally supporting the validity of the initial AHP results.

4.4. Future Prospects

4.4.1. Conducting New Surveys

Future surveys should increase the sample size and avoid relying solely on the Wenjuanxing platform. Extensive street surveys should be conducted repeatedly. Prior to filling out the questionnaire, it is advisable to clearly explain the existing problems of current products to the respondents, and then compare the new findings with those of this study.

4.4.2. Developing Practical Products

Research should not remain confined to the design phase. Future efforts will focus on product manufacturing and user trials to test and validate practical effectiveness [50].

4.4.3. Generalizability and Formalization of the Methodology

Furthermore, the design framework (the “LLM-AHP → SD-LLM → Manual Refinement” framework in Figure 1) can be formalized as a modular Input–Process–Output system. This system enhances the generalizability of the proposed methodology and can be applied to other product design scenarios.

The proposed methodology consists of two core functional modules. The requirement quantification module (LLM-AHP) transforms unstructured, multi-source data (interviews, literature, patents, and websites) into a structured, weighted hierarchy. Mathematically, this process maps a set of qualitative inputs $I_{\mathit{qu}alitative}$ to a set of quantitative weights $W_{\mathit{quantitative}}$, where ${\displaystyle \sum W_{\mathrm{i}}}=1$. This logic is domain-agnostic and can be applied to other product design scenarios requiring the prioritization of numerous user requirements. The generative iteration module (SD-LLM) establishes a closed-loop iterative system. In this system, the LLM serves as a “virtual advisory expert”, providing various improvement methods and suggestions based on the deficiencies identified by humans in the generated images, thereby guiding the SD model for regeneration. This “generation-deficiency identification-suggestion-regeneration” mechanism is applicable to other product design scenarios that require continuous improvement in image generation quality.

The framework can be extended to other hardware design scenarios by substituting the input data while maintaining the processing logic. If the framework is applied to designing smart educational toys for children, the data sources would shift to “smart educational toys” and “learning devices”. The LLM-AHP would likely prioritize “educational value” and “vision health” over “coercive control.” Subsequently, guided by these new weights, the SD-LLM module would generate design schemes focusing on smart educational devices and vision protection devices.

Therefore, the proposed methodology is not limited to anti-addiction devices but offers a robust, reusable workflow for designing smart hardware where user requirements are complex and subjective.

5. Conclusions

(1)

The LLM-AHP method addressed the issues of ambiguous requirements and significant subjective bias. It quantified these requirements and ranked them based on their weights. There were a total of 24 requirements, but their weight values varied significantly. The weights were generally divided into five intervals: 1–2 (very strong weight), 3–6 (strong weight), 7–11 (moderate weight), 12–16 (weak weight), and 17–24 (very weak weight). The top two requirements were as follows: “Coercive management and control for smartphone independent of self-discipline” (22.28%), and “Minimize human intervention, verbal persuasion, and family conflicts” (14.89%). The next four items were as follows: “Emergency use and coercive covering screen immediately after use” (9.94%), “Cannot be circumvented or deceived” (9.68%), “Emergency call” (9.47%), and “Coercive covering screen at the scheduled time” (7.02%). From the 7th rank, the value decreased to less than 5%, and from the 17th rank, it dropped to under 1%. This indicated that the key requirements concentrated in the top six items.

(2)

The SD-LLM method inspired the generation of novel schemes and resolved the issues identified by designers. The key design features such as the non-removable case, the non-standard internal triangular screws for anti-circumvention, the emergency call area, and the roller shutter to cover the screen were all derived from SD-LLM. However, traditional manual design remained indispensable for the creativity and design of complex components. The mechanism of the triple-verified contact detection device was quite complicated; it was designed by human designers.

(3)

The entire design process that included multi-source data collection, quantitative analysis, scheme iteration, and detailed manual refinement. It was reproducible and could be applied to the design of other smart products for children.

(4)

Limitations and future directions: The questionnaire sample size needs to be increased; the study should consider incorporating experts from diverse fields; future research would include user testing and reliability studies.