AI-Powered Prompt Engineering for Education 4.0: Transforming Digital Resources into Engaging Learning Experiences

1. Introduction

The rise of Education 4.0 has brought forward a vision of learning that is flexible, personalised, and intertwined with emerging technologies (World Economic Forum, 2023). Within this framework, the integration of AI into learning systems has been profoundly reshaping the dynamics of teaching and learning in digital environments. AI’s potential to automate, adapt, and personalise the educational process has been widely acknowledged (Castro et al., 2024), not only in higher education but across multiple academic levels (Bonfield et al., 2020). This transformation is not merely technological or superficial; it represents a paradigmatic shift in how we teach, learn, and make pedagogical decisions. Recent studies point to a significant growth in research focused on applying AI algorithms to support and guide students throughout their learning journey, with particular emphasis on Intelligent Tutoring Systems (ITS), personalised recommender engines, and adaptive learning (Alabi, 2025; Almufarreh & Arshad, 2023; An et al., 2023; Imran et al., 2024; Vergara et al., 2024; Zhou et al., 2025).

From a technical perspective, several AI techniques have been explored for these purposes, including Machine Learning (ML), Deep Learning (DL), and Natural Language Processing (NLP), among others(Eltahir & Babiker, 2024). These approaches enable, for instance, the prediction of learning difficulties based on interaction patterns, the generation of personalised study plans, and the provision of immediate and context-aware feedback. In online learning environments, AI-based systems are often integrated with learning analytics and educational data mining tools (Alabi, 2025; Imran et al., 2024; Vergara et al., 2024; Zhou et al., 2025), which collect and analyse large volumes of educational data to continuously enhance the teaching and learning experience. The study by (Vieriu & Petrea, 2025) highlights the effectiveness of these technologies in improving academic performance and student engagement, reporting statistically significant gains in groups using intelligent platforms compared to traditional methods, with strong correlations between AI use and success in formal assessments. Nevertheless, the application of AI in e-learning is not free from criticism or challenges, many of which are ethical or pedagogical (Dol & Jawandhiya, 2024; Eltahir & Babiker, 2024; Han et al., 2025). The opacity of algorithmic models raises concerns about automated decision-making and the potential for discriminatory bias. The use of sensitive personal data, often without informed consent, represents another controversial issue, especially in the context of growing concerns over digital privacy and data protection. Research such as that of (An et al., 2023; Eltahir & Babiker, 2024; Vergara et al., 2024) draws attention to the risks of technological dependency, loss of student autonomy, and the dehumanisation of the teaching-learning process (Dol & Jawandhiya, 2024).

At the same time, there is an ongoing debate surrounding the balance between automation and human intervention. While AI may serve as a powerful tool for identifying patterns and suggesting pedagogical strategies, it does not replace the role of the teacher as a critical mediator, promoter of reflection, and guardian of educational ethics. International regulatory frameworks are beginning to address this technological advancement by proposing guidelines for the responsible use of AI in education, including principles of fairness, explainability, inclusion, and algorithmic transparency. Students' perceptions of these technologies are also ambivalent. The study by (Eltahir & Babiker, 2024) shows that, although students recognise the potential of AI to personalise learning and improve study efficiency, they also express legitimate concerns regarding surveillance, data misuse, and the loss of control over their learning trajectories. In school settings, there is also evidence of limited literacy around how AI systems function, which may hinder critical and informed use. Given this context, it is essential to develop a better understanding of which AI techniques are being used to personalise learning, how these technologies have demonstrated effectiveness in improving educational outcomes, and what risks or limitations remain unresolved. Accordingly, the first part of this study conducts a systematic literature review using the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) protocol (Page et al., 2021) to map the current state of the art and provide a comprehensive, research-informed overview of AI applications in e-learning contexts. The PRISMA methodology offers a framework for conducting and reporting systematic reviews, designed to increase transparency, reproducibility, and rigour by guiding researchers through a standardised process of identifying, screening, assessing eligibility, and including studies. A central element of PRISMA is the use of a flow diagram that documents how records are retrieved from databases and how studies are progressively excluded or retained according to predefined criteria. By identifying practices, trends, gaps, and dilemmas, this work aims to contribute to a more comprehensive understanding on the ethical, technical, and pedagogical integration of AI in digital education and to propose a technique that addresses some of these existing gaps.

By identifying practices and trends and examining the importance of explicit or implicit prompts in various systems, the aim is, in a second phase, to present a practical proposal designed to facilitate the learning process through the integration of digital educational resources, delivered within e-learning environments, with an embedded prompt. The approach is structured to operate in alignment with the theoretical foundations underpinning ERS, ensuring that both pedagogical principles and the capabilities of the intelligent system guide the learning process.

The article is structured as follows: (i) Section 2 presents the research method supported by a systematic review methodology (PRISMA); (ii) Section 3 presents a practical demonstration of the integration of a prompt within digital learning resources; (iii) Section 4, the conclusions and suggestions for the implementation of future work are discussed.

2. Related Works

The use of artificial intelligence in the context of e-learning has gained increasing relevance, particularly in areas such as personalized learning, individualized monitoring, and educational content recommendations. ML methods—such as convolutional neural networks, recommender systems, and ITS are applied to recognise patterns in student behaviour and to adapt learning pathways to their specific needs. The scientific literature has explored these approaches, with a focus on their impact on academic performance, motivation, and student engagement in digital learning environments. This systematic review follows the PRISMA methodology (Page et al., 2021), and the topics: research questions; inclusion criteria; research strategy; results; data extraction, data analysis discussion, were included:

To conduct the research, we formulated research questions focused on the application of artificial intelligence in personalised learning within digital education environments. These questions aim to propose solutions to the identified problem.

2.1. Research Questions

The research questions we used are:

• Question 1: What artificial intelligence techniques have been applied to personalise students' learning pathways?
• Question 2: What is the reported effectiveness of these intelligent systems in improving students’ academic performance or engagement?
• Question 3: What are the limitations, risks, or criticisms associated with the use of AI for individualised monitoring in e-learning?

2.2. Inclusion Criteria

Inclusion criteria refer to the key characteristics of the target population that researchers will use to answer the study question. The inclusion criteria defined for our study are as follows:

• Criterion 1: Studies from 2023 to 2025.
• Criterion 2: Studies written in English and with full text available.
• Criterion 3: Studies that apply clearly identified artificial intelligence techniques in educational contexts.
• Criterion 4: Studies conducted in e-learning or digital education environments, at any educational level.
• Criterion 5: Studies that report outcomes related to academic performance, student engagement, motivation, or that discuss ethical, pedagogical, or technical limitations of AI-based learning systems.

2.3. Research Strategy

We searched for relevant articles using the ACM Digital Library (Acm digital library, 2025) and Scopus (Scopus, 2025) databases. The search terms we used ("artificial intelligence" OR "intelligent tutoring systems") AND ("e-learning" OR "digital education") AND ("personalised learning " OR "academic performance" OR "student engagement" OR "motivation" OR "ethical issues" OR "student autonomy" OR "limitations of AI"). We conducted our search between June and July 2025.

2.4. Results

After applying criterion 1, we identified a total of 166 scientific studies, comprising 49 from ACM Digital and 117 from Scopus, as presented in Figure 1. We conducted a comprehensive analysis of these studies, applying criteria 2, 3, 4, and 5, which resulted in a full-text analysis of the 99 remaining studies. Based on criteria 3, 4, and 5, the most relevant articles were selected, resulting in 54 articles being included in the review.

2.5. Data Extraction

Data were extracted from all the identified studies in the following format: Study, Educational Prompt Use, AI Techniques, Objective, Platform/Software, Education Level and Limitation Type. The table presented in Appendix A identifies the most essential characteristics of the selected studies. During the data extraction process, each study was determined individually. The AI techniques used were recorded to map technological trends and assess their effectiveness.

The role of Educational Prompt Use was systematically examined and added as a separate analytical category. The reviewed studies revealed three distinct situations: explicit use of prompts, typically in systems powered by chatbots or large language models, where learner instructions directly shape the AI’s output, implicit use of prompts, where student interactions such as responses, behavioural data, or feedback function as triggers without being explicitly framed as prompts and studies with no prompt-based interaction, relying instead on passive data collection such as biometric or clickstream analysis. This categorisation highlights the transversal role of prompts across AI-supported education and provides an additional lens through which to interpret how interaction design influences pedagogical effectiveness, personalisation, and learner engagement.

The objective of each application was described to provide context for the reported results. Platforms or software were specified to understand the technological environment in which the solution was implemented and how AI was integrated into these contexts. The education level was documented to analyse the suitability of the application for different age groups and educational settings, as well as to identify underexplored areas. Limitations were categorised to highlight the main barriers to implementation, covering technical, ethical, pedagogical, or contextual issues, thus enabling the identification of risks and the formulation of recommendations for future research.

The categorisation of artificial intelligence techniques adopted in this review was derived from taxonomies established by the ACM Computing Classification System (ACM, 2012). We grouped the techniques into the categories most frequently applied in educational contexts: machine learning (including deep learning and reinforcement learning), natural language processing (including language models, chatbots, and dialogue systems), recommendation systems, and computer vision (e.g., emotion recognition, engagement monitoring). Studies employing hybrid or multimodal approaches were identified separately when they combined methods from more than one category. Although no formal risk-of-bias tool was applied, a qualitative appraisal of methodological soundness was conducted, and the absence of a structured scoring system is acknowledged as a limitation.

2.6. Data Analysis and Discussion

The analysis of the 54 studies included in this systematic review, conducted according to the PRISMA protocol, reveals a growing trend in the adoption of AI for personalising digital learning between 2023 and 2025. This increase follows the accelerated digital transformation in education, which has become more evident after the pandemic, and reflects an increasing technological maturity in the approaches applied. Scientific production peaked in 2024, with 21 studies, followed by 2025 with 17 studies, and 2023 with 16 studies. This pattern suggests not only the evolution of AI tools but also the consolidation of specific methodologies and use cases. The presence in 2024 may be linked to the popularisation of language models, such as ChatGPT, and the growing integration of intelligent systems into teaching platforms.

The most common AI techniques, as illustrated in Figure 2, are ML and NLP. Machine learning appears in 52% of the studies, most frequently applied to tasks such as predicting academic performance and detecting behavioural patterns. NLP is present in almost 29% of the papers and has been increasingly enhanced by large language models, such as ChatGPT, enabling applications in intelligent tutoring, automated feedback, and personalised interaction. Recommender systems, although less frequent at 6% of the studies, play a significant role in adapting content and suggesting resources aligned with learners’ profiles. Hybrid or multimodal AI approaches account for 8% of the cases, combining techniques to provide richer insights into learning processes. Computer vision is the least frequently used approach, appearing in 5% of the studies, typically in emotion recognition, engagement monitoring, and activity detection. This distribution highlights the dominance of machine learning approaches, the growing relevance of NLP enhanced by generative AI, and the emergence of multimodal systems, while also indicating that vision-based applications and recommender systems remain relatively underexplored in educational contexts.

The analysis of the selected studies, illustrated in Figure 3, indicates that the role of Educational Prompt Use can be differentiated into three main situations. Explicit use of prompts is the most frequent, appearing in 54% of the studies, particularly in systems based on chatbots or large language models, where learner instructions directly determine the system’s output and guide personalised feedback. Implicit use of prompts is less common, identified in 18% of the works, where student inputs such as quiz responses, behavioural data, or interaction logs act as triggers for adaptive mechanisms, even though they are not explicitly described as prompts. Finally, 28% of the studies reveal no prompt-based interaction, focusing instead on passive data collection methods such as facial recognition, attention monitoring, or biometric analysis. This distribution demonstrates that prompting is not limited to generative AI but is instead a transversal mechanism across different applications of AI in education. Whether explicit or implicit, prompts function as key drivers of personalisation and adaptive feedback, shaping the interaction between learners and systems.

The analysis of the platforms/software reported in the reviewed studies is illustrated in Figure 4. The platform/Software referred to in the studies shows a clear predominance of e-learning and online platforms, which represent 20% of the cases. This confirms their role as the primary environment for integrating AI into educational contexts. ChatGPT is the second most cited tool, with 12%, reflecting the rapid uptake of large language models for tutoring, personalised interaction, and automated feedback.

AI-enabled platforms and mobile AI-based platforms each account for 11%, underlining the growing importance of adaptive systems and the expansion of mobile learning solutions. Other platforms appear less frequently but still demonstrate the diversity of technological environments. These include AI-assisted platforms, AI-powered chatbots, custom AI systems, generative AI tools, IoT-enabled platforms, MOOCs, smart education platforms, and virtual learning environments, each with a 5% share. More specialised or emerging tools are mentioned only occasionally, such as learning management systems, visual programming environments, and virtual reality platforms, each accounting for 2%.

Overall, the distribution suggests that while mainstream e-learning platforms continue to be the backbone of AI adoption in education, there is an evident diversification of approaches. The significant presence of ChatGPT and other generative AI tools highlights a shift towards intelligent conversational agents. At the same time, the emergence of mobile, IoT, and immersive platforms indicates exploratory yet promising avenues for future development.

The education levels presented in Figure 5, based on the reviewed studies, indicate that research on AI in education is concentrated in specific segments. Higher education accounts for 60% of the studies, reflecting the strong interest in applying AI tools to universities and colleges, where digital infrastructures and access to large datasets make implementation more feasible. K-12 education, which spans primary through secondary schooling, accounts for 22% of the workforce, often focusing on adaptive learning platforms, AI-assisted tutoring, and engagement monitoring to support personalised instruction. Other levels of education appear less frequently but remain relevant. Corporate training, lifelong learning, and professional training each account for 6% of the studies. These works typically explore domain-specific training, continuous professional development, or workplace learning enhanced by adaptive and immersive technologies. This distribution demonstrates a clear predominance of higher education as the main testing ground for AI in education, with K-12 as the second most studied area. By contrast, corporate, professional, and lifelong learning remain comparatively underexplored, despite their potential to expand AI-driven personalisation and skill development beyond formal education. The availability of digital infrastructure and the relative ease of data collection and processing in academic institutions likely explain the intense focus on higher education.

The distribution of limitation types, illustrated in Figure 6, is based on classifying each article according to the most relevant limitation reported. However, several articles mentioned more than one limitation, and some even addressed all three categories. Technical limitations are the most frequently reported, appearing in 41% of the studies. These often relate to high computational demands, integration difficulties with existing platforms, limited scalability, and dependency on high-quality datasets. Pedagogical constraints account for 39%, including challenges such as aligning AI-generated learning materials with curriculum standards, maintaining student engagement, and ensuring that AI complements rather than replaces human-led teaching practices. Ethical concerns are identified in 20% of the works, focusing on issues such as data privacy, bias in AI models, transparency of automated decisions, and the potential for reinforcing educational inequalities. This breakdown highlights that technical barriers are currently the top challenge to integrating AI in education. Pedagogical and ethical considerations remain essential. All three categories must be addressed in a balanced way to ensure trustworthy and impactful adoption.

Overall, the review of the 54 articles reveals that AI in education is dominated by established techniques such as ML, DL, and NLP, complemented by emerging generative AI applications and specialised algorithms for complex tasks. Higher education emerges as the primary field of application, followed by K12, which spans from primary to secondary schooling, with far fewer studies in vocational, lifelong, and cross-level contexts. In terms of limitations, most studies face pedagogical challenges, with technical barriers and ethical concerns also present but less frequent. While classification in this review focuses on the most relevant limitation for each article, many studies acknowledge multiple constraints, and a notable number discuss all three types. This suggests that the successful integration of AI in education will require not only technological advancements but also pedagogical innovation and strong ethical frameworks to ensure sustainable and equitable adoption.

Based on the information obtained, the following answer to the research questions is presented:

Question 1: The analysis of the studies reveals a broad spectrum of artificial intelligence techniques implemented to enhance the personalisation of students’ learning trajectories in digital education contexts (Ilić et al., 2023). Personalisation is implemented through several families of methods. Classic recommendation approaches are presented in (Amin et al., 2023; Huang et al., 2023; Zhang, 2025), which utilise collaborative filtering, content-based filtering, data mining, and learning analytics to recommend resources and sequence activities. Adaptive modelling is central to (Baba et al., 2024; Castro et al., 2024; Gligorea et al., 2023; Halkiopoulos & Gkintoni, 2024; Modak et al., 2023) where learner state is inferred and content difficulty is adjusted dynamically. Advanced control and sequential decision methods are employed in (Bagunaid et al., 2024; Sharif & Uckelmann, 2024) which learn intervention policies from multimodal traces. Predictive modelling that feeds adaptation is reported in (Gámez-Granados et al., 2023; Q. Huang & Chen, 2024; Zhen et al., 2023). Natural language technologies act as personal tutors and guides in (Alrayes et al., 2024; Ayoubi, 2024; Bellot et al., 2025; Dahri et al., 2025), as well as in simulation-based settings (Stampfl et al., 2024), where dialogue systems and LLMs personalise help, explanations, and practice. Context-aware and affect-aware personalisation are discussed in (Alshaya, 2025; Dhananjaya et al., 2024; Mutawa & Sruthi, 2024; Yang et al., 2025). Systems-level enablers of personalisation are discussed in (Haque et al., 2024; Koukaras et al., 2025; Singh et al., 2025; Wang & Sun, 2025; Wang & Liu, 2025), which address edge computing, networking, and platform integration that make low-latency personalised adaptation feasible. Finally, domain-specific implementations (An et al., 2023; Hu & Jin, 2024; Miranda & Vegliante, 2025; Yang, 2024; Yong, 2024; Zheng, 2024)

Question 2: Most of the reviewed works reported some measure of system effectiveness, whether through academic performance indicators, engagement metrics, completion rates, or motivation scales (Babu & Moorthy, 2024; Villegas-Ch et al., 2024). Multiple empirical studies report positive effects on achievement, engagement, or self-efficacy. The study (Jafarian & Kramer, 2025) shows gains in reading outcomes and motivation when speech technologies structure practice and provide feedback. The research (Huang et al., 2023), links targeted recommendations to increased engagement and better assessment scores. In mobile and conversational contexts, (Abdulla et al., 2024; Dahri et al., 2025; Zhu et al., 2025), report improved self-efficacy, faster task progression, or higher marks. The study (Yang et al., 2025) associates’ context-aware practice with better vocabulary retention, while (Haque et al., 2024), reports performance benefits from continuous monitoring and tailored support. Early warning and prediction studies, such as (Bagunaid et al., 2024; Huang & Chen, 2024; Zhen et al., 2023). The Document improved predictive accuracy, which enables timely intervention, a proximal driver of improved outcomes. Affective and behavioural sensing in (Mutawa & Sruthi, 2024; Zhu et al., 2023) enhances detection of disengagement and supports adaptive responses that maintain participation. In creative and simulation-based learning, (Stampfl et al., 2024; Zeng et al., 2025; Zheng, 2024), report higher engagement, more authentic practice, and perceived gains in higher-order skills. Literature reviews and system-level papers, including (Amin et al., 2023; Castro et al., 2024; Gligorea et al., 2023; Halkiopoulos & Gkintoni, 2024), synthesise evidence that adaptive sequencing, timely feedback, and targeted recommendations are associated with improved learning processes and outcomes across diverse settings.

Question 3: Limitations span pedagogical, technical, and ethical domains, and many articles acknowledge more than one. Privacy and autonomy concerns are prominent in monitoring-focused works such as (Hossen & Uddin, 2023; Mandia et al., 2024; Mutawa & Sruthi, 2024; Rahman et al., 2024; R. Zhu et al., 2023), which involve the collection of fine-grained behavioural or biometric data and raise questions about consent, transparency, and potential misuse. Technical barriers frequently cited include data quality and scalability limitations in (Gámez-Granados et al., 2023), computational and integration costs in (Bagunaid et al., 2024; Q. Huang & Chen, 2024; Sharif & Uckelmann, 2024), and speech or text recognition errors in (Elbourhamy, 2024; Zhen et al., 2023). Infrastructure and security dependencies are emphasised by (Haque et al., 2024; Yong, 2024; Zhen et al., 2023). Pedagogical critiques include potential over-reliance on AI reported in (Abdulla et al., 2025; Alrayes et al., 2025; Bellot et al., 2025; Ilieva et al., 2023; Suh et al., 2025) as well as variable effectiveness across learners, documented in (An et al., 2023; Y. Yang, 2024; Zeng et al., 2025). Broader ethical and methodological concerns are synthesised in (Ali et al., 2025; Rahe & Maalej, 2025; G. Wang & Sun, 2025), which discusses academic integrity, bias, and equitable access (Mourabit et al., 2025). The capability limits of current models are clearly outlined in (Mendonça, 2024), which highlights reasoning and diagram-understanding constraints. Usability and adoption risks are highlighted in (Alsanousi et al., 2023), and design trade-offs are discussed in (Ovtšarenko & Safiulina, 2025). Context and generalisability limitations are highlighted by (Martín-Núñez et al., 2023) and by the dependence on infrastructure or devices in (Baba et al., 2024). Finally, language and cultural fit issues are highlighted in (Alshaya, 2025; Miranda & Vegliante, 2025), where translation quality and affect recognition may be misaligned for diverse cohorts. These findings highlight the need for explainable, privacy-conscious, and pedagogically aligned AI tools, underscoring the importance of continuous evaluation in real-world deployments.

Overall, the cross-analysis of these findings indicates that, while AI, particularly machine learning, NLP, and recommender systems, has shown potential to improve student performance and engagement, technical, ethical, and pedagogical limitations remain significant barriers. This underscores the need to develop explainable, transparent, and pedagogically aligned systems to maximise the positive impact and minimise risks in the use of AI for personalised learning in digital education environments.

Thus, this analysis demonstrates that while AI techniques, particularly recommendation systems, have a measurable positive impact on student performance and engagement, these solutions still face considerable challenges. Overcoming these challenges requires more transparent, integrated and user-centred approaches, ensuring that technical or ethical barriers do not compromise pedagogical value. The reported results indicate significant gains in academic performance, student motivation, and engagement. In several cases, increases in completion rates, improvements in assessment results, and greater perceived relevance of recommended content were observed. The personalisation enabled by these systems allows for the optimisation of study time and the promotion of more effective learning trajectories, contributing to more meaningful and engaging educational experiences.

The systematic analysis in Section 2 identified a consistent body of evidence on the potential and limitations of using artificial intelligence in digital learning contexts. While various techniques, ranging from recommendation systems to adaptive platforms, demonstrate significant improvements in student performance and motivation, gaps remain concerning transparency, pedagogical integration, and the explainability of these systems. However, the review also highlights that few studies evaluate the pedagogical design of prompts or systematically measure their impact on learning outcomes. This gap reinforces the need for further research to establish prompt use and subsequent design as a critical element in the effective integration of AI into teaching and learning. These findings not only offer a critical framework for understanding current developments but also guide the creation of innovative solutions to address these limitations. In response, Section 3 introduces a specific methodological approach: integrating digital resources with embedded prompts into AI agents. This work aims to develop a methodology and practical application designed to optimise the learning process through the integration of digital educational resources in e-learning environments and the use of prompts embedded within AI agents.

3. Practical Demonstration – Embedded Prompt

This section presents a practical demonstration of a methodology devised to optimise the learning process through the integration of digital educational resources, made available in e-learning environments, with a prompt embedded within an AI agent, in alignment with the theoretical foundations that underpin ERS.

The proposal aims to generate AI-guided digital learning resources, designed to transform static educational materials into dynamic and personalised learning experiences (Marzano, 2025). It leverages the capabilities of large language models (LLMs) to develop interactive study guides from conventional didactic content, actively fostering learner engagement and autonomy (Baídoo-Anu & Ansah, 2023). The approach is implemented through a structured process comprising three distinct phases, as illustrated in Figure 7.

Phase 1 – Digital Learning Resources: This initial phase commences with pre-existing digital learning resources, such as text documents (.docx), presentations (.pptx), spreadsheets (.xlsx), or documents in PDF format. These resources, frequently hosted on e-learning platforms and integrated within VLEs, constitute the knowledge base wherein the pedagogical content resides.

Phase 2 – Insertion of the Learning Guide Prompt: The core of the framework is based on the integration of a structured learning guide prompt. This combines a set of metadata and directives, including, among other things, the definition of the role to be assumed by the AI (e.g., 'expert tutor'), the pedagogical objective, and the intended output format.

Phase 3 – AI-Guided Digital Learning Resource: The Digital Learning Resource and the prompt are submitted to an AI agent (e.g., Gemini, ChatGPT, Claude). The agent processes the instructions and generates a new digital artefact: the guided learning resource.

Thus, the ascent of LLMs represents a paradigm shift in human-computer interaction, displacing the focus from explicit programming to natural language instruction (Brown et al., 2020; Liu et al., 2021; Sahoo et al., 2025). At the centre of this new interaction modality lies the prompt: the textual input provided to the model to request the generation of a response. The conception of prompts has evolved beyond its initial definition as a mere question or command. Presently, its elaboration is recognised as a discipline, prompt engineering, dedicated to the construction of complex textual artefacts that guide, constrain, and optimise the model's behaviour for specific tasks (Liu et al., 2021; Sahoo et al., 2025). Formally, a prompt P can be represented as a sequence of tokens P = (t1, t2,…, tk), concatenated with a user input, to maximise the probability of generating the desired output.

The scientific evolution of prompts can be segmented into several phases. Initially, their potential was demonstrated through in-context learning techniques, such as zero-shot and few-shot prompting (Brown et al., 2020; Kojima et al., 2023), wherein the model executes tasks without additional training, relying solely on a description or a few examples included within the prompt itself. This approach revealed the capacity of LLMs to generalise from direct instructions, but also their sensitivity to the formulation and format of the input. A decisive advance was the introduction of techniques to enhance the models' reasoning processes. The most notable of these is Chain-of-Thought (CoT) prompting, which instructs the model to generate a sequence of intermediate steps before the final answer (Wei et al., 2023), thereby improving performance on arithmetic, logical, and symbolic reasoning tasks. From this foundation, variants such as Self-Consistency have emerged, which generate multiple reasoning pathways and select the most coherent answer (X. Wang et al., 2023), demonstrating that the prompt's structure influences not only the final response but also the computational process itself. The study in question was limited to the analysis of four specific file formats, examining the feasibility and efficacy of the covert prompt insertion technique within each, with the PDF format serving as an example. Thus, considering a practical example that illustrates the three phases in Figure 8, this involves transforming a Python exercise sheet in PDF format into an interactive tutorial session. In this session, the AI agent not only provides solutions but also engages the student, inquiring about their preferred starting point, demonstrating an understanding of the original document's structure (e.g., the number of questions), and offering flexible learning pathways.

The elaboration of effective prompts for LLMs also paves the way for their utilisation as virtual tutors, capable of promoting autonomous, reflective, and rigorous learning (Liu et al., 2021; White et al., 2023). The most advanced practices treat the prompt as a quasi-formal specification of the desired behaviour (Sahoo et al., 2025), based on four key principles:

• Role Delegation: assigning the model an explicit persona (e.g., 'Act as an expert tutor…') to guide its behaviour, tone, and knowledge (Liu et al., 2021; White et al., 2023).
• Rich Contextualisation and Task Delimitation: providing detailed context and structural delimiters (e.g., <context>...</context>) to segment relevant information (White et al., 2023).
• Explicit and Structured Instructions: avoiding ambiguities by decomposing instructions into clear or conditional steps (Sahoo et al., 2025).
• Definition of Constraints and Output Format: specifying rules (<rules>) and formats (<output_format>) to ensure predictability and integration with other systems (Greshake et al., 2023; Zou et al., 2023).

Beyond semantic design, there are technical approaches for the covert insertion of prompts, which are relevant to educational and security contexts. These include the concealment of text within documents (.docx, .pdf, .pptx, .xlsx) through chromatic formatting and structural positioning, thereby maintaining functional integrity while rendering the content invisible to the end-user (Greshake et al., 2023; Zou et al., 2023). Thus, the analysis of complex prompts that integrate a pedagogical persona, a formative context, guided reasoning, and rigorous operational constraints proves fundamental to understanding how these components materialise into engineered artefacts designed to maximise the pedagogical efficacy of LLM-based systems (Liu et al., 2021; White et al., 2023).

3.1. Prompt Components

One of the central elements in prompt engineering is the delegation of a persona (role delegation). Defining a specific role not only serves to anchor the model's behaviour but also to guide the discursive tone and the attributes that characterise the interaction. By assuming an explicit identity and objectives, the model adopts a coherent perspective that is aligned with the user's expectations and the task's context. In the case under analysis, the chosen persona is that of a specialist teacher across various fields of knowledge, to whom specific characteristics can be attributed, as illustrated. This framework aims to ensure that all responses emanate from the perspective of an experienced educator, geared towards supporting teaching and learning processes.

Thus, the <role> component not only assigns a clear identity to the model but also establishes solid guiding principles for more effective pedagogical interactions. Another structuring element is the explicit definition of the target audience (target age group). This aspect plays a determinant role in the appropriateness of the language, the choice of examples, and the depth of the generated content, allowing the model's response to be calibrated to optimise relevance, clarity, and pedagogical impact. In this case, the defined audience corresponds to adult learners (18 years and older), encompassing university students, working professionals, and individuals engaged in lifelong learning. This delimitation ensures that the model adjusts the level of complexity, as well as the type of references and applications, to respond effectively to the needs of this audience.

The <target_age_group> component thus functions as a mechanism for communicative adaptation, aligning the discourse with the expectations and needs of the target audience. The pedagogical efficacy of a prompt also depends on how the feedback is structured.

The <feedback_level> component specifies the style and formative function of the input within the teaching and learning process. In the example under analysis, the feedback level is formative and personalised, geared towards stimulating the student's cognitive autonomy. Instead of correcting answers directly, it promotes reflection, the understanding of errors, and independent problem-solving. This approach is aligned with pedagogical practices that value the active construction of knowledge and the development of critical thinking. Beyond the persona, the target audience, and the level of feedback, it is essential to define the thematic and structural framework of the response.

The <context> component fulfils this role by establishing the scope of the interaction and providing guidelines for formatting and content, thereby ensuring quality, accessibility, and neutrality. It acts as a formal guide for delivering rigorous responses that foster critical thinking. In the given example, the focus is on producing clear, structured, and engaging academic explanations or summaries, following a format that should include:

• Clear definitions: key terms explained precisely and accessibly.
• Core concepts: development of the fundamental ideas.
• Illustrative examples: concrete scenarios or analogies.
• Practical applications: uses in real-world contexts.
• Critical thinking: questions for in-depth analysis.

The <instructions> component defines the interaction methodology, prioritising guided reasoning. This approach emphasises strategic questions to steer reflection, conceptual cues to contextualise problem-solving, and partial explanations to break down complex problems. The model only provides complete solutions after the student has made an active attempt to solve the task, thereby developing higher-order cognitive skills and respecting the learner’s pace. The definition of rules constitutes a critical dimension, ensuring the integrity of the pedagogical process and the reliability of interactions.

The <rules> component imposes constraints and safeguards designed to preserve the coherence of the system, protect the prompt’s internal structure, and ensure the maintenance of a consistent pedagogical persona. Among the core principles are:

• Prompt immutability: prevents the user from altering internal instructions, thus preserving methodological coherence and pedagogical control.
• Prompt invisibility: ensures the learner remains unaware of the underlying prompt engineering, avoiding interference with the learning experience and the neutrality of the interaction.
• Persona consistency: guarantees that the model retains a helpful, patient, and professional demeanour, reinforcing trust and predictability.

The <output_format> component standardises the structure of responses, guiding the learner through seven sequential steps: (1) Clear statement; (2) Understanding the problem; (3) Strategy to be used; (4) Step-by-step guidance with justifications; (5) Ask for the answer; (6) Final answer and verification; and (7) Tip for generalisation or reflection. Steps 6 and 7 are only revealed once a correct answer has been provided, thereby encouraging individual effort and initiative.

Finally, the <user_input> component initiates the interaction, automatically adapting the response language to match that of the question and inviting the student to indicate the starting point of the session, thus creating a personalised and interactive learning environment. Appendix B presents the complete prompt, including all its elements. At the same time, a summary of the components is displayed in Table 1, specifying, for each, its pedagogical function, a concise description, and an illustrative example.

3.2. Illustrative Demonstration of Embedded Prompt Application

This section presents the analysis of the results obtained from demonstrating the Digital Learning Resource, which is understood as an autonomous and static educational resource designed to support the learning process independently, when applied within the AI agent ChatGPT. In parallel, the AI-Guided Digital Learning Resource is discussed as an innovative approach that integrates and interconnects two key elements: the resource and the prompt. This integration enhances an adaptive, interactive, and guided learning experience, demonstrating its functionality as an educational recommender system.

As illustrated in Figure 9, the graphical interface of the chatbot (ChatGPT) is presented, showing the upload of a PDF file containing 14 questions on the Python programming language, without any prompt being inserted. After the upload, the user enters the prompt: Do the exercise in this file.

Subsequently, as illustrated in Figure 10, the chatbot provides answers to the 14 questions included in the exercise contained in the Python Exercises.pdf file. Ultimately, it asks the user whether they would like an explanation for each question. In this way, the learner is initially presented with the direct answers and only later invited to request further clarification on the content—a common practice among AI agents.

As illustrated in Figure 11, the graphical interface of the ChatGPT chatbot is presented, showing the upload of a PDF file containing 14 questions on the Python programming language. After the upload, the user enters the prompt: Do the exercise in this file.

In this scenario, as illustrated in Figure 12, the Digital Learning Resource has the prompt embedded, which leads the AI agent to follow the established instructions rigorously. The interaction is thus initiated as defined in the <user_input> component: Begin by asking the student: "Which exercise, or topic would you like to start working on today?"

The agent fully interprets the content of the file and interacts with the student, asking them to indicate the question they wish to address or, within the pedagogical content, in this case, Python programming, the topic they want to review. For example, should the student choose to deepen their learning regarding Question 5, and in accordance with the prompt definition, the response format is specified in the <output_format> component, which structures the responses and guides the student through seven sequential steps. The following section provides a detailed analysis of each step.

The first step, “1. Clear Statement”, as illustrated in Figure 13, clarifies the question by identifying the character used to define a comment in Python. Four answer options are provided. The content is structured clearly, with visual emphasis on the question number and title, followed by the question text and the available options.

Figure 14 presents the next component, “2. Understanding the Problem”, which aims to guide reasoning before answering the previous question. The text proposes two reflective questions, organised in a bulleted list format, encouraging the reader to recall the essential concepts regarding comments in Python before selecting the correct answer.

The component “3. Strategy To Be Used”, illustrated in Figure 15, outlines the recommended approach for accurately addressing the question concerning the comment character in Python. The text advises recalling the concept of comments in Python, reflecting on practical examples, and identifying the correct symbol employed for commenting.

Figure 16 presents the component “4. Step-By-Step Guidance with Justifications”, which provides detailed explanations for each response option in the exercise.

The image in Figure 17 presents the component “5. Ask For The Answer”, in which the reader is prompted to indicate, based on the previously developed reasoning, which option they consider correct. The text also inquires whether the user wishes to respond before receiving confirmation. In the bottom-right corner, the indication “option a” appears, possibly representing the user's selected answer.

Figure 18 illustrates the interaction between the AI agent and the user in the context of correcting an error made during the resolution of an exercise. A pedagogical approach is employed, aimed at gradually clarifying content through sequential explanations tailored to the level of understanding demonstrated by the student. This process is accompanied by positive reinforcement strategies, which foster a supportive learning environment and encourage cognitive autonomy.

Finally, Figure 19 illustrates the moment at which the system presents the correct answer, in accordance with the principles outlined in the “6. Final Answer and Verification” component. This process includes not only the confirmation of the solution but also an explanatory synthesis of the methodological pathway adopted to obtain it, thereby contributing to the consolidation of learning. Subsequently, the application of the guidelines from the “7. Tip for Generalization or Reflection” component is observed. These guidelines involve formulating a concluding reflection, an extension question, or a proposal for a practical application to foster knowledge transfer and stimulate students' critical thinking.

In summary, interaction with the student allows for the personalisation of the learning experience by offering content, activities, or resources adapted to the needs, preferences, and performance of students, fundamental characteristics of an educational recommender system that operates within VLEs (Drachsler & Kalz, 2016; Lu et al., 2015)..

Various LLMs have developed interactive strategies that help students actively build their knowledge. These strategies focus on the learning process rather than just giving answers (Kasneci et al., 2023). They also reduce the need for specific prompts to guide the AI's teaching behaviour.

For example, in the current context, ChatGPT offers the 'Study and Learn' mode, which acts as a tutor, guiding the user step-by-step with examples, analogies, and exercises, applying active learning, scaffolding, and adaptive feedback to reinforce comprehension, retention, and autonomy (Brown et al., 2020; OpenAI, 2024; Yaseen et al., 2025), Meanwhile, Gemini, in its 'Guided Learning' mode, follows an incremental approach by breaking down problems, exploring errors as opportunities, and adjusting the pace and complexity to the user, with a focus on critical thinking and enduring comprehension (Google DeepMind, 2024). Claude, in its 'Explanatory' style, leverages clear communication, conceptual decomposition, and practical examples, adjusting its language and technical level to the user's profile to promote critical reflection and scientific rigour (Anthropic., 2024).

Thus, an evolution is observed from simple information providers to active educational partners, personalising learning pathways and promoting transferable skills. This advancement faces the challenge of balancing autonomy, rigour, and transparency. It is therefore argued that this transformation should also be incorporated into the development of digital educational resources, fostering strategic alignment with the evolution of critical thinking and digitally mediated study (Nye et al., 2014).

In brief, the practical demonstration in this section illustrates how the structured embedding of prompts in AI agents can transform static digital resources into adaptive learning experiences aligned with pedagogical principles and students' needs. This methodological proposal is not isolated; it directly addresses the gaps identified in the systematic review in Section 2, especially regarding the insufficient integration of technological personalisation with pedagogical principles, and the need for greater transparency and explainability in AI-based educational systems. This approach bridges the gap between current research and the exploration of innovative solutions, paving the way for the final reflections and development perspectives discussed in the next section.

4. Conclusions

The systematic review, conducted in accordance with the PRISMA methodology, confirms that the application of artificial intelligence (AI) techniques to the personalisation of learning in e-learning environments constitutes a rapidly expanding field, with globally consistent results. Among the most widely used approaches are machine learning, natural language processing, neural networks, and, more recently, large language models (LLMs) and reinforcement learning. In this context, recommender systems play a central role due to their capacity to combine performance data, student preferences, and navigation patterns to suggest individualised learning pathways.

Despite these advances, the literature shows that the use of prompts in education remains under-theorised and has largely been operationalised as a mere technical trigger, rather than as an intentional pedagogical instrument. This gap underscores the need to integrate AI’s technical potential with robust pedagogical principles, ensuring that personalisation is not only effective but also educationally meaningful.

Nonetheless, certain limitations of this review must be acknowledged: potential publication and selection biases, restriction to English-language sources, and the absence of a formal risk-of-bias assessment tool. While these factors do not invalidate the findings, they warrant caution in their interpretation.

In response to the identified gaps, a prompt engineering architecture was developed, consisting of seven interdependent components (persona, target audience, feedback, contextual framing, reasoning instructions, operational rules, and output format). This proposal illustrates how static content can be transformed into interactive experiences, with potential to foster autonomy, metacognition, and critical thinking. However, its effectiveness remains to be empirically demonstrated, particularly regarding the robustness of the persona model, adaptation to different cultural contexts, and the assessment of metacognitive gains.

Accordingly, future research should focus on: (i) comparatively testing the methodology across different AI agents; (ii) optimising prompts in relation to emerging ethical challenges in education; (iii) integrating and refining the proposal within e-learning systems; and (iv) validating the approach in real classroom contexts, particularly in teacher education. Only through such an applied research programme will it be possible to transform this conceptual proposal into practical, reproducible, and pedagogically grounded evidence, thereby contributing to more personalised, meaningful, and autonomous learning pathways.