NoTeeline: Supporting Real-Time, Personalized Notetaking with LLM-Enhanced Micronotes
Faria Huq, Abdus Samee, David Chuan-En Lin, Alice Xiaodi Tang, Jeffrey P. Bigham · 2025 · Proceedings of the 30th International Conference on Intelligent User Interfaces (IUI '25) · doi:10.1145/3708359.3712086
Summary
This paper introduces NoTeeline, an interactive notetaking tool that uses LLMs to expand user-written "micronotes" — brief shorthand jottings like "plastic pol. ->" or "RNNs are unrolled l to r or opp" — into full-fledged notes that maintain the user's personal writing style. The system addresses a fundamental tension in notetaking: manually written notes miss crucial details due to the fast pace of content, while automatically generated notes lack personalisation and discourage active engagement. NoTeeline's approach keeps users as active participants who decide what to capture and how to frame it, while offloading the mechanical task of expanding abbreviated thoughts into complete sentences. The system is built as a React.js web app with a Chakra UI interface inspired by the Cornell notetaking method, featuring three panels: Notes (for writing and expanding micronotes), Cues (auto-generated review questions), and Summary (personalised takeaway). The LLM pipeline (GPT-4-Turbo) uses video transcripts at the micronote's timestamp for context and a small collection of the user's previous notes from an onboarding session (three short clips with corresponding full notes) for writing style personalisation. Five atomic operations are supported: micronote expansion, note organisation by theme, note modification, cue question generation, and summary generation.
Key findings
In a within-subjects study with 12 participants, NoTeeline demonstrated significant improvements across multiple dimensions. Users wrote 47.0% less text (mean 32.11 vs 60.64 words per note) and completed notes 43.9% faster (mean 15.59 vs 27.79 seconds per note) while actually capturing more notes overall (mean 18.75 vs 14.75). Video disruption decreased dramatically — users paused or rewound only 0.56 times with NoTeeline versus 4.94 times with baseline. The expanded notes achieved 93.2% factual correctness (measured by HHEM hallucination evaluation) and 8.33% improvement in writing style consistency (measured by chi-squared distance between the generated text and the user's manual writing). NASA-TLX scores showed significantly lower Mental Demand (p=0.0123), Effort (p=0.0024), and Frustration (p=0.0047). SUS usability scored 80.97 (grade A). The system correctly handled diverse micronote patterns including abbreviations ("l to r" -> "left to right"), symbols ("->"), question marks for uncertainty, and personal shorthand. All 12 participants wished to continue using NoTeeline. The needfinding study revealed that users' primary concern with AI-assisted tools was lack of agency — they wanted "a tool for me... to help me get a better understanding, not a tool for me to tell me what I should know."
Relevance
NoTeeline has direct accessibility implications, as noted in the paper's societal impact statement: it can make notetaking more accessible for individuals who struggle with language fluency (non-native speakers) and physical disabilities that make typing difficult. The micronote paradigm is particularly powerful for accessibility — users with motor impairments or slower typing speeds need only capture the essence of an idea in a few characters, and the system fills in the rest. The approach also reduces cognitive load, benefiting people with cognitive or learning disabilities who find simultaneous listening and writing challenging. More broadly, NoTeeline exemplifies a design principle that runs through much of Bigham's recent work: AI should amplify rather than override human capabilities. The user remains the author — deciding what to note and providing the seed content — while the AI handles mechanical expansion. This "implicit guidance" model, where the micronote implicitly directs the AI without requiring explicit prompting, represents a transferable pattern for accessible AI tools that need to minimise additional user effort while maintaining agency.
Tags: large language models · writing assistance · personalization · notetaking · cognitive load · education · video accessibility · user agency