PlanGlow: Personalized Study Planner

• PlanGlow is a personalized study planning system that uses LLMs to dynamically create, critique, and refine custom learning plans aligned with educational theories.
• It emphasizes transparency through explainability features, such as contextual rationales and real-time resource validation via external APIs.
• The system enhances user controllability by allowing interactive refinements and iterative feedback that tailor plans to individual learning needs.

PlanGlow is a personalized paper planning system that leverages LLMs to generate, critique, and refine self-directed learning plans with a focus on explainability and user controllability. The system is structured to address the limitations of prior LLM-based educational tools, particularly with respect to transparency of recommendations, control over plan content, and the mitigation of LLM hallucination. PlanGlow combines rigorous integration of educational theory, user-centered interaction models, and real-time resource validation to produce paper plans that are academically robust and tailored to individual learner needs (Chun et al., 16 Apr 2025).

1. System Architecture and Workflow

PlanGlow operates as a web-based application with a React.js and CSS front end and a Python FastAPI back end. Core functionalities interface with the OpenAI GPT-4 API, configured via a multi-phase “chain-of-thought” prompting strategy. This process is explicitly detailed in the system’s documentation and implementation figures.

The workflow proceeds as follows:

1. User Input: Collection of paper preferences via a structured form. Inputs include subject area, learning goals, background knowledge, paper timeline, and daily availability.
2. Initial Plan Generation: The LLM generates a draft plan with:
   • A layered, hierarchical structure (week-by-week overview plus daily tasks).
   • Objectives and suggested learning materials.
3. Critique Phase: The draft is then critiqued by the LLM against predefined educational theories and self-directed learning frameworks, such as Bloom’s Taxonomy, Vygotsky’s Zone of Proximal Development, and Knowles’ principles of adult learning.
4. Improvement Phase: Using feedback from the critique, the system re-prompts the LLM—adjusting prompting parameters such as temperature and top$_p$—to output a refined plan that incorporates hierarchical objectives, resource justifications, and context-sensitive explanations.

This iterative, model-in-the-loop process is architected to overcome the opaqueness and rigidity that characterize many LLM-based learning tools.

2. Explainability and Transparency Mechanisms

PlanGlow explicitly integrates explainability at both the interface and plan-content levels. Users are presented with contextual reasoning for weekly and daily paper tasks, rationales for choice of resources, and hierarchical explanations that reference established educational theories when relevant.

Transparency features include:

• Display of the educational or cognitive rationale for each recommended activity or resource.
• Real-time validation of external resource links (e.g., YouTube) using the YouTube Data API v3, with visibility into data such as view and like metrics.
• Progressive disclosure interfaces allowing users to access varying depths of explanation, from high-level summaries to detailed justifications.

These mechanisms directly address LLM opacity and hallucination by grounding each plan recommendation in explicit, inspectable logic and external validation.

3. User Controllability and Interaction Design

Controllability—defined as the user's ability to tailor, refine, and iterate on their paper plan—is engineered into PlanGlow at multiple stages:

• Structured input forms and in-line editing enable rapid modification of paper parameters (duration, objectives, knowledge level).
• A chat-style interface allows re-prompting for further refinements, requests for alternative paper resources, or clarification on plan content.
• Interactive selection and replacement of paper materials; users can swap out recommended resources with system-vetted alternatives.
• All user-driven changes are propagated through the LLM pipeline, ensuring that downstream critique and improvement stages respect updated user context.

These features were identified as critical during formative studies and are codified in the system’s design requirements (D3: enhanced controllability).

4. Integration of Educational Theory

The critique and improvement phases of PlanGlow’s pipeline are framed by explicit references to human learning theory. The LLM is prompted to align recommendations with:

• Bloom's Taxonomy: Structuring objectives from knowledge recall to higher-order skills.
• Vygotsky's Zone of Proximal Development: Calibrating content complexity to learner capability.
• Knowles' Adult Learning Principles: Emphasizing learner autonomy, goal relevance, and experiential linkage.

This grounding enables, for instance, the system to justify a task as “targeting Bloom’s application level” or to reference the zone of proximal development when sequencing concepts. Pedagogical soundness is thereby made explicit, both to the end user and in system-internal quality checks.

5. Technical and Experimental Evaluation

A mixed-methods approach was adopted for PlanGlow’s development, comprising:

• Formative Surveys and Interviews: 28 survey responses and 10 in-depth interviews (plus an educational expert) which informed the requirements for the system’s explainability, controllability, material reliability, and usability.
• Within-Subject Experiment: A controlled paper (N=24) compared PlanGlow against two baseline systems: a GPT-4o-based tool and Khan Academy’s Khanmigo. Power analysis confirmed statistical sensitivity (effect size $d=0.8$; $\alpha$ corrected to $0.05/3$ by Bonferroni adjustment).
• Quantitative Evaluation: Participants constructed plans with each system and rated usability, explainability, and controllability via 22 Likert-scale items. ANOVA with Bonferroni post-hoc comparison revealed significant PlanGlow advantages ($p<.01$ on several explainability and controllability metrics).
• Expert Assessment: Two educational specialists rated paper plans on clarity of objectives, timeline realism, and pedagogical soundness using a 5-point matrix.

Evaluation results demonstrated:

• Comparable raw plan quality across systems, but PlanGlow’s superiority in integration of controllability and explainability features.
• PlanGlow was selected as preferred by 83.3% of participants.
• Expert review confirmed that PlanGlow’s plans provided clearer objectives, timelines, and pedagogical consistency.

6. Prompt Engineering and Parameterization

The chain-of-thought prompting sequence is detailed, and key LLM parameters are tuned for each stage to optimize factuality and minimize repetitive or unsupported content. For example:

$$\begin{aligned} &\text{Initial generation:} \quad \text{Temperature} = 0.2,\, \text{Top}_{p} = 0.6,\, \text{Frequency\_penalty} = 0.2,\, \text{Presence\_penalty} = 0.1 \ &\text{Refinement phase:} \quad \text{Temperature} = 0.0,\, \text{Top}_{p} = 0.8,\, \text{etc.} \end{aligned}$$

This LLM parameterization is critical for plan consistency and factual accuracy. The chain-of-thought methodology (see Figure 1 in the primary source) ensures that critique and improvement operate with explicit reference to input parameters and established learning science.

7. Comparative Positioning and Significance

PlanGlow’s primary distinguishing features are its integrated, user-facing explainability and its multi-level controllability—attributes sparsely addressed in prior LLM-based planning systems. Unlike baseline systems such as GPT-4o-based planners or Khanmigo, PlanGlow provides context-sensitive justifications and allows interactive, user-directed modification of both procedural and material plan components.

Evaluation studies suggest that PlanGlow’s approach:

• Resolves core issues of LLM opacity, resource hallucination, and user agency.
• Enables rapid, iterative customization of paper plans aligned to established pedagogy.
• Bridges the gap between LLM generative flexibility and rigor in educational practice.

A plausible implication is that such systems, if deployed widely, may reshape both self-directed paper and broader applications of explainable AI in education, particularly where transparency and user control are paramount.

---

Table: Key Comparison Dimensions of PlanGlow and Baseline Systems

Feature	PlanGlow	Baseline Systems (GPT-4o, Khanmigo)
Explainability	Hierarchical, theory-based	Minimal or ad-hoc
Controllability	Multi-step, interactive	Limited, mostly initial-stage only
Resource Validation	Real-time, API-based	Little or no validation
Plan Structuring	Layered, objective-driven	Typically flat or less contextual
Expert Endorsement	Positive	Mixed or unassessed

PlanGlow thus exemplifies a rigorous, experimentally validated approach to explainable and controllable AI-driven educational planning. Its architecture, methodology, and empirical findings provide a reference implementation for future research and development in AI-based learning environments.