PersonaMail: Learning and Adapting Personal Communication Preferences for Context-Aware Email Writing

TL;DR

This work developed PersonaMail, a system that addresses gaps in LLM-assisted writing through structured communication factor exploration, granular editing controls, and adaptive reuse of successful strategies, and contributes design implications for AI-assisted communication systems that prioritize interpersonal nuance over generic text generation.

Abstract

LLM-assisted writing has seen rapid adoption in interpersonal communication, yet current systems often fail to capture the subtle tones essential for effectiveness. Email writing exemplifies this challenge: effective messages require careful alignment with intent, relationship, and context beyond mere fluency. Through formative studies, we identified three key challenges: articulating nuanced communicative intent, making modifications at multiple levels of granularity, and reusing effective tone strategies across messages. We developed PersonaMail, a system that addresses these gaps through structured communication factor exploration, granular editing controls, and adaptive reuse of successful strategies. Our evaluation compared PersonaMail against standard LLM interfaces, and showed improved efficiency in both immediate and repeated use, alongside higher user satisfaction. We contribute design implications for AI-assisted communication systems that prioritize interpersonal nuance over generic text generation.

Figures (4)

• Figure 1: System overview of PersonaMail showing the integrated architecture for tone articulation and email composition. (A) In the task prompt stage, users simply describe the email's core purpose without specifying tone. (B) The Factor Exploration Panel scaffolds users' reflection on research-based communicative factors through (B1) structured quick selection and (B2) open-ended elaboration. (C) The text-edit interface enables granular tone adjustments through communicative-unit-based editing. The email body is segmented into labeled components such as Opening Salutation, Justification, and Closing Pleasantry (C1). When users select a segment, its associated communicative intents (C2)—such as Explanation Specificity, Formality Level, and Emphasis on Regret. Users can preview and apply alternative strategic variations (e.g., shifting from vague reason to full disclosure), observing immediate text-level updates. The Adaptive Stylebook's Quick Fix interface (C3) surfaces personalized revision patterns learned from user edits. When users select text, the system presents alternative phrasings that reflect previously observed stylistic tendencies (e.g., softening tone or increasing privacy), transforming past edits into adaptive, reusable communication instruments. (D) Persona-Situation Anchors allow users to save successful communication strategies. The system enables creating anchors based on recipient relationships (Persona, D1) or communication contexts (Situation, D2), with AI-generated descriptive names that users can refine to match their reuse intentions. (E) In future writing stages, users can apply saved anchors to pre-configure the system's communicative factors, generating an initial draft that mirrors their preferred tone expectations with minimum effort.
• Figure 2: The PersonaMail System Pipeline. The workflow illustrates the information flow from initialization to adaptive learning. (1) Context & Initialization: User intent is captured via the Factor Exploration Panel and retrieval and adaptation of existing Persona/Situation Anchors. (2) Drafting: The LLM generates a base draft. (3) Analysis & Structure: The system decomposes text into Communicative Units and links them to identified Intents. (4) Refinement & Utilities: Users refine the draft using Intent-Driven Modifications, QuickFix, and other AI-based tools. (5) Learning & Feedback: The system captures manual edits to update the Adaptive Stylebook and allows users to save new Anchors, which feed back as the Adaptive Cycle for future tasks.
• Figure 3: Quantitative evaluation results comparing PersonaMail with the baseline (Gemini+Gmail). PersonaMail significantly outperformed the baseline across email quality, overall efficiency, and cognitive load. (A) First draft quality: PersonaMail's initial drafts were rated 34.8% higher in quality than the baseline's (p < 0.001). (B) Revised draft quality: After revision, PersonaMail emails maintained significantly higher quality ratings (p = 0.002). (C) Time efficiency: While PersonaMail required more upfront articulation time in the first use, its adaptive features led to a 42.0% reduction in total task time upon reuse compared to the baseline, and a 50% reduction compared to its first usage (p < 0.001). (D) Cognitive load: PersonaMail reduced users' cognitive burden by 24.5% compared to the baseline, as measured by Raw-TLX scores (p < 0.001). Error bars represent standard deviation. ***p < 0.001, **p < 0.01.
• Figure 4: Representative systems used in our formative study spanning different approaches to LLM-assisted email composition.