ReelsEd: AI-Driven Microlearning Reels

• ReelsEd is a web-based system that automates conversion of long lecture videos into structured microlearning reels using large language models.
• It integrates transcript extraction, GPT-4 summarization, modular video editing, and instructor oversight to ensure pedagogical fidelity.
• Empirical results show improved quiz performance and task efficiency, establishing ReelsEd as a valuable tool in AI-driven educational media.

ReelsEd is a web-based system that automates the conversion of long-form lecture videos into structured, short-form microlearning reels using LLMs. Its methodology centers on identifying and extracting pedagogically salient moments from instructor-authored material, then assembling concise video segments that maintain clarity, instructional integrity, and high learner engagement. The architecture integrates transcript processing, generative AI for summarization, modular video editing, and instructor oversight to produce educational media optimized for efficiency and usability. Empirical evaluation demonstrates statistically significant improvements in learning outcomes and user trust, establishing ReelsEd as a reference implementation for AI-driven educational video generation (Stavrinou et al., 7 Sep 2025).

1. System Architecture and Workflow

ReelsEd operates as a multi-component pipeline designed to generate short-form educational reels from lengthy lecture videos. The process comprises:

• Transcript Extraction: Source videos are downloaded, and full transcripts are fetched using the YouTube transcript API.
• Key Moment Identification: Extracted transcripts are processed with GPT-4 via staged prompts. The initial GPT-4 prompt scans for timestamps associated with instructional “key moments” that align with the video’s learning goals.
• Summarization and Labeling: For each key moment, a subsequent GPT-4 invocation generates a succinct summary and explicit label, designed to preserve the original instructional intent and clarity.
• Video Trimming and Assembly: Using MoviePy, the system programmatically trims the video at the identified timestamps and assembles these segments into a set of polished, short-form reels.
• Instructor Review and Editing: Teachers have the option to inspect and refine the generated reels, ensuring content alignment with intended pedagogical outcomes.

The system is constructed on Python (Django backend), PostgreSQL (data layer), NGINX (reverse proxy), and employs Docker for environment consistency. A complementary React Native mobile application affords instructors parameter control over reel segmentation (e.g., number, duration).

2. Generative AI Summarization Methodology

Central to ReelsEd is its use of GPT-4 for multi-stage transcript summarization. Rather than inventing new content, the model is prompted to identify and distill key instructional moments based on learning objectives extracted from the video transcript.

• GPT-4 receives a transcript and associated metadata, identifies segments that correspond to core concepts, and generates short summaries with clear semantic labels.
• The summarization pipeline is explicitly designed to preserve instructor-authored material, reinforcing the system’s educational authenticity.
• Instructor-editable outputs allow human oversight to correct, supplement, or rephrase AI-generated content.

This design ensures that the reels maintain the pedagogical fidelity of the original material, which is critical for both instructional value and learner trust.

3. User Study and Quantitative Results

A between-subjects laboratory paper evaluated ReelsEd against traditional long-form lecture video delivery, focusing on learning efficacy, engagement, cognitive load, and trust.

• Study Population: 62 university students were randomly assigned to either the control (long-form) or experimental (ReelsEd-generated reels) condition.
• Assessment Protocol: Participants viewed the assigned content, completed comprehension quizzes, and responded to standardized questionnaires: UEQ (user experience), perceived learning efficacy, NASA TLX (cognitive load), and a trust instrument for the experimental group.
• Results:
  • Quiz Performance: The reels group scored higher (μ ≈ 93.85% ± 7.89) than the long-form group (μ ≈ 79.72% ± 16.98), U = 736.50, p = 0.0001.
  • Task Efficiency: The reels group completed quizzes faster (328.77s) than the long-form group (446.23s), U = 219.00, p = 0.0002.
  • Engagement: The modular, bite-sized format was consistently rated as easier to process and more engaging.
  • Cognitive Load: No increase in perceived cognitive load was reported for the reel condition.
  • Trust: Participants expressed high trust in the system’s dependability, accuracy, and integrity, with high agreement on positive statements and low agreement on negative statements regarding deception.

The paper utilized robust statistical analysis (Mann–Whitney U test) and open-ended thematic coding for qualitative insights.

4. Usability and Pedagogical Alignment

ReelsEd emphasizes design features that support usability, agency, and pedagogical alignment:

• Interface: A stepwise UI guides instructors and learners through segmentation, review, and playback.
• Authenticity: Instructor-authored material is maintained; the LLM’s role is limited to summarization.
• Agency: Adjustable reel parameters and navigability allow for tailored and self-directed learning experiences.

This approach aligns with evidence-based principles in cognitive load theory and modular microlearning, optimizing both educational rigor and learner autonomy.

5. Implications for Educational Media and Future Directions

Findings from ReelsEd suggest substantial opportunities for integrating generative AI into scalable educational media systems:

• Microlearning Optimization: Automated summarization and segmentation support modular, focused instruction that is compatible with diverse educational settings.
• Pedagogical Integrity: The workflow maintains the instructor’s intent and voice, reducing risk of content distortion.
• Scalability and Efficiency: Containerized deployment and modular architecture facilitate adaptation across platforms and institutional contexts.

A plausible implication is that future educational tools might leverage similar architectures for efficient microcontent generation, coupled with enhanced human-AI collaboration and oversight to preserve instructional quality.

6. Relationship to Related Work

ReelsEd’s approach is conceptually distinct from systems targeting narrative beat extraction for media retargeting (Nguyen et al., 26 Mar 2024) or video-level vision-language understanding with LLM adapters (Reza et al., 7 Apr 2025). Unlike Scrolly2Reel, which repurposes interactive graphics for social platforms, and REEF, which optimizes token relevance for multimodal video tasks, ReelsEd’s pipeline is optimized for instructor-controlled microlearning with empirical pedagogical evaluation. These differences highlight varied pathways for AI-driven video summarization and retargeting, situating ReelsEd within the broader research context of generative educational media.

7. Summary Table: System Components

Component	Function	Technology Stack
Transcript Extractor	Fetches YouTube transcript	YouTube API, Python
LLM Summarizer	Identifies and summarizes key moments	GPT-4
Video Editor	Trims and assembles reels	MoviePy, Python
Web/Mobile Interface	Instructor control and review	Django, React Native
Data Backend	Stores video and user data	PostgreSQL, NGINX

All components are modular and containerized for cross-platform deployment. This technical breakdown reinforces the system’s scalability and reproducibility for academic and institutional adoption.