Paper deep dive

PuppetChat: Fostering Intimate Communication through Bidirectional Actions and Micronarratives

Emma Jiren Wang, Siying Hu, Zhicong Lu

Year: 2026Venue: arXiv preprintArea: cs.HCType: PreprintEmbeddings: 111

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%

Last extracted: 7/20/2026, 7:51:40 PM

Summary

The paper introduces PuppetChat, a dyadic messaging prototype designed to foster intimate communication through bidirectional actions and personalized micronarratives. It addresses the limitations of current instant messaging tools by using a reciprocity-aware recommender to suggest responsive actions and generating micronarratives from user stories to ground interactions in personal history. A 10-day field study with 11 dyads demonstrated that this approach enhances social presence, supports expressive self-disclosure, and sustains continuity and shared memories.

Entities (6)

PuppetChat · software-system · 98%Micronarrative · concept · 95%Reciprocity-aware recommender · algorithmcomponent · 92%Social Presence · metricoutcome · 92%Field Study · research-method · 90%Instant Messaging · technology-domain · 88%

Relation Signals (5)

PuppetChat → generates → Micronarrative

confidence 95% · generates personalized micronarratives from user stories

PuppetChat → uses → Reciprocity-aware recommender

confidence 95% · PuppetChat uses a reciprocity aware recommender to encourage responsive actions

Field Study → evaluates → PuppetChat

confidence 92% · Our 10-day field study with 11 dyads... revealed that this approach enhanced social presence

PuppetChat → enhances → Social Presence

confidence 90% · revealed that this approach enhanced social presence

Instant Messaging → ischallengedby → PuppetChat

confidence 85% · today's tools often dilute care... To address this challenge, we present PuppetChat

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:As a primary channel for sustaining modern intimate relationships, instant messaging facilitates frequent connection across distances. However, today's tools often dilute care; they favor single tap reactions and vague emojis that do not support two way action responses, do not preserve the feeling that the exchange keeps going without breaking, and are weakly tied to who we are and what we share. To address this challenge, we present PuppetChat, a dyadic messaging prototype that restores this expressive depth through embodied interaction. PuppetChat uses a reciprocity aware recommender to encourage responsive actions and generates personalized micronarratives from user stories to ground interactions in personal history. Our 10-day field study with 11 dyads of close partners or friends revealed that this approach enhanced social presence, supported more expressive self disclosure, and sustained continuity and shared memories.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

110,180 characters extracted from source content.

Expand or collapse full text

figures/ -nc-nd PuppetChat: Fostering Intimate Communication through Bidirectional Actions and Micronarratives Emma Jiren Wang jiren@vt.edu 0009-0003-4413-9944 Virginia TechBlacksburgVirginiaUSA , Siying Hu siyinghu@cityu.edu.hk Department of Computer ScienceCity University of Hong KongHong KongSAR and Zhicong Lu zlu6@gmu.edu Department of Computer ScienceGeorge Mason UniversityFairfaxVirginiaUSA (2026) Abstract. As a primary channel for sustaining modern intimate relationships, instant messaging facilitates frequent connection across distances. However, today’s tools often dilute care; they favor single tap reactions and vague emojis that do not support two way action responses, do not preserve the feeling that the exchange keeps going without breaking, and are weakly tied to who we are and what we share. To address this challenge, we present PuppetChat, a dyadic messaging prototype that restores this expressive depth through embodied interaction. PuppetChat uses a reciprocity aware recommender to encourage responsive actions and generates personalized micronarratives from user stories to ground interactions in personal history. Our 10-day field study with 11 dyads of close partners or friends revealed that this approach enhanced social presence, supported more expressive self disclosure, and sustained continuity and shared memories. Intimate Communication, Social Presence, Micronarrative, Large Language Models, Continuity, Expressivity †journalyear: 2026†copyright: c†conference: Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems; April 13–17, 2026; Barcelona, Spain†booktitle: Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI ’26), April 13–17, 2026, Barcelona, Spain†doi: 10.1145/3772318.3790685†isbn: 979-8-4007-2278-3/2026/04†ccs: Human-centered computing Collaborative and social computing systems and tools Figure 1. Conceptual vision and interface implementation of the expressive messaging system. (Top) A conceptual story sequence illustrating the types of emotional support and reciprocal physical interactions the system aims to facilitate between users. (Bottom) Interface mockups demonstrating the actual translation of these concepts into mobile UI features. The center lists four key Design Goals (DG1–DG4). Arrows indicate how specific UI components, such as the recommended avatar actions under ”Physicalness” (implementing DG1 & DG2) and the saveable interaction history under ”Memories” (implementing DG4), directly address these design goals. Conceptual vision and interface implementation of the expressive messaging system. (Top) A conceptual story sequence illustrating the types of emotional support and reciprocal physical interactions the system aims to facilitate between users. (Bottom) Interface mockups demonstrating the actual translation of these concepts into mobile UI features. The center lists four key Design Goals (DG1–DG4). Arrows indicate how specific UI components—such as the recommended avatar actions under ”Physicalness” (implementing DG1 and DG2) and the saveable interaction history under ”Memories” (implementing DG4)—directly address these design goals. 1. Introduction Instant messaging (IM) has become a primary medium for sustaining relationships across distance, yet in this setting intimate expression is often diluted by the logic of convenience (Nardi et al., 2000; Verheijen, 2013; Cho et al., 2017): when people truly need to be understood, comforted, or “do something together,” communication is compressed into a handful of icons or a block of text that takes significant effort to compose, either too light and easy to misread or too heavy and hard to write. Conversations frequently resolve into non-invitational and terminal feedback (Woerner et al., 2007; Schwarz, 2011), such as long pressing a message to attach a thumbs-up reaction, which neither conveys care promptly nor offers a way to link the present moment with prior shared experience (Clark et al., 2019; Groffman et al., 2010). Prior work shows that feeling “we are here together” is crucial for maintaining relational closeness, emotional support, and shared meaning in mediated communication (Wijesinghe et al., 2025). Without this sense of presence, communication may appear efficient yet still fail to nurture connection. To enhance relatedness under spatial separation, prior work highlights a set of complementary strategies, namely Awareness, Expressivity, Physicalness, Joint Action, and Memories (Hassenzahl et al., 2012; Falbo and Peplau, 1980; Overall et al., 2009). Taken together, they argue that intimate communication should not prioritize convenience alone; it should gently make partners feel ”seen,” enable emotional expression that is personalized yet leaves room for interpretation, restage care through appropriate embodied cues (Ichino et al., 2022; Starke et al., 2024), transform replies from endpoints into invitations to the next move to sustain continuity (Hassenzahl et al., 2012; Adler et al., 2010), and accumulate moments into a revisitable shared timeline. Guided by this lens, we take these elements as design levers to reconfigure everyday IM for intimacy. We present PuppetChat, a dyadic messaging prototype that operationalizes these elements in everyday chat. To support awareness, PuppetChat offers low friction action-plus-micronarrative exchanges, where an action is a short expressive puppet animation (e.g., hug, wipe-tears) and a micronarrative is a brief caption style text that contextualizes the action, allowing a partner’s momentary state to surface without demanding attention. For expressivity, it generates editable, story driven micronarratives anchored in user provided personal stories, allowing meanings to be tailored beyond generic emojis. To approximate physicalness, PuppetChat uses embodied puppet avatars capable of performing 42 actions (e.g., hug, wipe-tears) that restage care rather than merely label it. To support joint action, the system uses an LLM based recommender that suggests actions that naturally complement the partner’s action (e.g., when one partner selects cry, the system may suggest wipe-tears). This turns individual actions into opportunities to co-perform. Finally, to cultivate memories, actions and micronarratives are preserved in the thread as lightweight traces that can be revisited, helping interactions accumulate into a shared timeline. To evaluate how PuppetChat supports intimate expression, joint action, and relational presence in everyday messaging,across a 10-day field study with 11 dyads of close partners (N=22), we found that reciprocal actions reliably increased social presence and a felt sense of co-performance even in a non-immersive, text based interface. Lightweight and responsive actions paired with micronarratives turned replies into sequences rather than one-off reactions, maintaining the flow of interaction instead of breaking it into disconnected turns. Editable and story driven micronarratives afforded precise yet face saving self disclosure and seeded playful co-construction of meaning (e.g., inside jokes and private riffs). Over time, the accumulation of actions and captions created revisitable traces that supported reflection on shared moments. Taken together, these effects lowered the friction of nuanced expression while strengthening reciprocity and continuity, which ultimately enhancing partners’ sense of relatedness in intimate relationships. Our research thus makes three primary contributions to the HCI community: • A novel interface paradigm for intimate communication that integrates embodied action with personalization micronarratives. We present PuppetChat, a system demonstrating how lightweight bidirectional avatar actions and AI generated micronarratives can support more reciprocal and narratively grounded dyadic chat. • Empirical insights into how reciprocal and embodied interactions foster connection in mediated contexts. Our field study identifies mechanisms by which users employ responsive actions for ”co-performance” and editable narratives for playful ”co-construction” of meaning, thereby strengthening relatedness in intimate relationships. • Design implications. We outline concrete directions for technologies that move beyond one directional reactions to foster deeper narrative depth and interactional continuity in digital relationships. 2. Related Work This work reviews prior studies in three key areas: (1) awareness and social presence, which investigate how peripheral and ambient cues foster a sense of ”being seen” and ”being together” across distance; (2) physicalness and joint action, which examine the use of physiological proxies and re-enacted gestures or rituals to recreate bodily co-presence within instant messaging contexts; and (3) expressivity and micronarrative, which explore enriched communicative forms ranging from minimal affective signals to dyad specific symbolic languages and short, revisable narratives that bridge lightweight cues and extended messages. 2.1. Awareness and Social Presence To evoke a sense of ”being seen” and ”being together” in intimate relationships, a large body of work employs peripheral, low intrusion cues that permeate everyday settings: rather than relying on explicit conversation, signals are embedded in spaces and objects so that information flows continuously in the background (Oh et al., 2018; Larsen and Johnson, 2017). Interfaces and artifacts in this line prioritize being felt at the periphery over commanding center stage attention (Stone et al., 2005), using light, implicit cues to cultivate social presence. For example, merely indicating that ”the other is there,” diffusely mapping everyday activity rhythms, or conveying affect through environmental/physiological hints (Sproull and Kiesler, 1986; Miyata and Osawa, 2024). Unlike message centric approaches, the value here lies in continuity without interruption: the partner’s existence and routines are sensed in a subtle, always on manner that folds into daily practice, instead of launching a new explicit exchange. Empirical studies show that such ”small yet continuous” cues can heighten state awareness and connectedness, and help counter over idealization in long distance contexts by supporting a more realistic mutual understanding (Hassenzahl et al., 2012; Wijesinghe et al., 2025; Ahmmed et al., 2024). Turning awareness into a positive experience, however, depends on matching intensity, frequency, and audience of disclosure to relationship stage and situation: cohabiting versus long distance, work versus public settings, all differ in tolerance for sound, ambience, and rhythm cues (Riordan and Kreuz, 2010; Ames et al., 2010). Consequently, prior work often stresses interpretable ambiguity and boundary control (Fattorini, 1968): leaving room for receiver interpretation to reduce feelings of surveillance, while allowing the sender to adjust visibility (to whom, when, at what granularity). At the same time, limitations emerge. Many awareness signals remain one way indicators weakly coupled to conversational turn taking (Hassenzahl et al., 2012; Bhandari and Bardzell, 2008). They increase being seen but do not necessarily invite the next move, and they rarely connect systematically to relationship narratives or shared memory (Roediger and Pyc, 2012). Taken together, peripheral awareness cues succeed at sustaining ”being seen,” yet they often stop at one directional indication, insufficiently integrated with IM’s turn taking mechanics and thus poor at inviting follow up interaction. They also face a trade off between attentional burden and intrusiveness, where signals that are too subtle are overlooked (Swallow and Jiang, 2013); too explicit and they disrupt the moment, and their contextual fit and disclosure granularity are frequently not tunable across relationship stages and public or private settings. In response, we embed awareness cues within the IM dialogue flow via low friction, peripherally perceivable action prompts (which can be sent action only for weak disclosure and minimal disruption) and use complementary response mechanisms to move from merely ”being seen” to being received, encouraging a next turn with symmetric or complementary actions. This reframes presence from a static indicator into respondable co-presence, while offering controllable intensity and visibility (e.g., ephemeral status versus more persistent options) so that salience and interactivity improve without sacrificing privacy boundaries. In this way, awareness becomes more than a faint background signal: it lightly activates reciprocity and a felt sense of togetherness, thereby more effectively supporting social presence. 2.2. Physicalness and Joint Action When partners are apart, the most palpable loss is the sense of bodily co-presence (Shi et al., 2024; Hanks, 2013). Prior work responds along two routes: one uses physiological proxies (e.g., heart rhythm, skin warmth, rhythmic vibration) to evoke the bodily side effects of proximity (Morgan, 1943); the other re-enacts gestures (e.g., holding, stroking, hugging) through wearables or actuated objects to stage experiences of being held or cared for (Liu, 2024; Zhu et al., 2025; Pettys-Baker et al., 2024; Tang et al., 2024). In parallel, research on joint action seeks to re-establish a rhythm of doing together (Podkosova et al., 2023), either by inventing new dyadic rituals (Fourie et al., 2022) (e.g., small coordinated activities across bedrooms or remote mini games) or by piggybacking on mundane routines (Homaeian et al., 2021) (e.g., synchronized sipping). The two lines share a goal: physicalness brings back the felt texture of care, while joint action recreates a shared tempo (Hassenzahl et al., 2012). Touch is a direct conduit for empathy and care, yet mediation introduces stubborn frictions (Hertenstein et al., 2009; Knierim et al., 2025). Immediacy is difficult: embraces and hand holds rely on prompt back and forth, whereas technical delays sap reciprocity (Andersen et al., 1979; Vasseleu, 1999). Context fit is fragile: what feels tender in private can appear conspicuous or awkward in public, pushing users to abandon visible devices (Liao et al., 2024); many systems therefore pivot from one-to-one haptics to symbolic or poetic evocations of care (Kianzad and Seifi, 2024; Venkatraj et al., 2025). On the joint-action side, the target is behavioral interdependence (Wageman, 2014; Rusbult and Van Lange, 2008; Kong et al., 2025; Ikeya et al., 2024), but many implementations settle into parallel play or light synchronization that couples weakly to IM turn taking and rarely invites a next move (Li et al., 2022). Selecting activities that both trigger mutual influence and embed smoothly in everyday life across settings remains a central design tension. In sum, mediated touch can evoke closeness but struggles with synchronous reciprocity and situational appropriateness; joint-action systems foster togetherness yet often depend on artificial rituals or weak interdependence. What is missing in IM contexts is a low friction, turn embedded means to restage care as action and to transform replies from endpoints into invitations, without demanding strict synchrony or heavy hardware. We address this by integrating embodied metaphors of care directly into reply structure and leveraging complementary actions (e.g., distressed move → soothing response) so that physicalness is enacted within the chat’s turn system and joint action emerges as a reciprocal rhythm. This keeps salience and intrusiveness tunable while foregrounding behavioral interdependence in everyday messaging. 2.3. Expressivity and Micronarrative Expressive communication in intimate contexts goes beyond merely ”saying how one feels”; it also uses encoded or enriched forms that make feelings visible, answerable, and even playful (Bruns et al., 2021; Gross and John, 1998; Hübler, 2011; Ståhl et al., 2005). Prior systems span two broad lines: one centers on minimal on–off signals (Panda et al., 2025)(e.g., a press, a squeeze mapped to light or vibration) to create ultra low friction affective touchpoints in both synchronous and asynchronous use (Roto et al., 2016); the other invites partners to co-create a dyad specific symbolic language (Griggio et al., 2021; Alsayahani, 2023), composing messages from shared symbols (Fay et al., 2018; Fowler, 1987; Simpson, 1990) whose meanings only the two of them understand. Both lines rely on interpretable ambiguity: the same cue can take on different meanings by context, preserving discretion in public while sustaining a sense of ”us” (Baxter and Bullis, 1986; Crawford and Hamidi, 2025; Andalibi et al., 2017). Relationship research underscores that emotional expression is necessary for maintenance (Wagener et al., 2025; Kim et al., 2023; Furman and Wehner, 1997; Keltner et al., 2019); suppressing it incurs cognitive costs (Keltner et al., 2019), and in long distance settings more frequent affective writing is linked to stability (Wang et al., 2023). Intimate exchange also carries a strong reciprocity expectation (Burgoon and Le Poire, 1993; Rosenthal, 1967): expressions call for timely replies, so tools must embed into daily rhythms (Li et al., 2018), lowering the threshold to speak and making it easy to be ”caught.” (Reid et al., 1996) Yet tensions persist. Minimal cues often collapse into terminal feedback (a tap that ends the exchange), symbol systems can miscalibrate tone or be misread, and people differ in expressing and recognizing emotion; gender, culture (Ale-Ebrahim et al., 2023; Griggio et al., 2019), and the handling of negative affect (argument, repair) are frequently under addressed (Jones et al., 2011). Between low friction but semantically thin signals and nuanced but effortful long text, IM lacks a context coupled, editable middle layer. Short, revisable micronarratives (Bizzocchi et al., 2014; Skeggs et al., 2025) address this gap: a line or two tethered lightly to the moment (and optionally to an action) that clarifies tone, marks intent, and preserves face while inviting a next move. Unlike fixed symbols, micronarratives emphasize rewritability and fit to context (Bizzocchi et al., 2014); unlike full messages, they remain lightweight and improvisable. In everyday IM, such units can shift expression from a one way display to a turn worthy contribution, providing fine grained semantic footholds for continued, reciprocal interaction. 3. Formative Study We adopted an iterative design process. As a first step, we conducted a small scale formative study using an early demo prototype of PuppetChat to explore users’ initial reactions, which informed the design of the more complete prototype developed afterward. 3.1. Participants We recruited 5 participants (3 women and 2 men) for this study. All participants were frequent instant messaging users who regularly engaged in dyadic chats to sustain their daily social and relational communication. None of them had prior exposure to our system. Participants reported using a range of IM platforms in their everyday life, and were familiar with common expressive resources such as emojis, stickers, and GIFs. 3.2. Study Procedure We used a prototype demonstration (Bødker and Grønbæk, 2020) method to gather early reactions to PuppetChat’s core interaction concepts. Each 30–40 minute session began with an informed consent and a brief introduction to the study objectives. Participants were then shown an early demo prototype designed to present only the minimal idea of using animated actions in chat. The prototype used an abstract purple avatar to avoid implying gender or identity and keep attention on the interaction concept. We included twelve actions that provided a small but expressive set of relational gestures, selected through early sketching and informed by prior work on expressive messaging(Nardi et al., 2000; Hübler, 2011; Ståhl et al., 2005). Each action GIF displayed a short action description on hover, enabling participants to confirm the intended meaning of the action during exploration. The interface allowed participants to enter a short text message, after which the system used an LLM to analyze the input and recommend four candidate actions (Fig. 3 A1). Selecting one of these options triggered the corresponding GIF animation. For comparison purposes, we also showed two sample animations (throw-heart and carry-heart) from the human-like girl avatar used in the finalized system and from a robot avatar, although these were presented only briefly at the end of the session rather than as part of the interactive prototype. The demo did not include personalized micronarratives, personal stories, a replay function, or complementary action recommendations based on a partner’s actions, allowing us to probe user expectations without introducing predetermined design directions. Participants were asked to (1) enter short messages, (2) review the recommended actions, and (3) trigger action animations using the purple avatar. After exploring these interactions, a semi-structured interview invited reflections on their expectations, preferences, and everyday IM practices. During the interview, participants were also shown the two additional avatar examples and asked for brief feedback. They further provided examples of how they currently use emojis, stickers, and GIFs, and discussed how animated actions might extend or disrupt these existing practices. 3.3. Analysis Interview notes and transcripts were analyzed using a lightweight thematic analysis approach(Braun and Clarke, 2019). Two authors independently reviewed the session notes, generated codes based on what appeared meaningful in the data, and refined these through discussion until agreement was reached. The aim of the analysis—given the exploratory nature of the formative study—was to surface notable user expectations and potential design opportunities. 3.4. Findings Our findings partly aligned with expectations grounded in prior literature. For example, users’ desire for reciprocal exchanges (F1) reflects well established observations about turn taking and mutual responsiveness in mediated communication (Hassenzahl et al., 2012; Ling, 2010). Similarly, the value of personalization and contextual cues (F2–F3) echoes prior work on intimacy and relational maintenance in messaging systems (Alsayahani, 2023). At the same time, the study surfaced several unanticipated needs, such as the desire for flexible persistence (F4), which were not explicitly highlighted in earlier research. These insights informed the design of the full prototype and extended the system beyond what prior work alone would have predicted. F1: Users sought reciprocal exchanges and clearer guidance for responses Participants emphasized that expressive actions should not be isolated, one-off tokens like emojis or stickers. Instead, they wanted actions to invite a partner’s response, creating a closed loop of interaction. Several participants described frustration when an action such as a ”hug” felt incomplete without a corresponding reply, noting that ”if the other person doesn’t respond, it feels unfinished.” To address this, they suggested that the system could highlight when a response was received or recommend complementary actions to make reciprocity easier—for example, pairing ”throw-heart” with ”catch-heart” or suggesting playful alternatives such as ”split-heart” or ”throw back the heart.” These ideas point to the importance of embedding reciprocal structures into action design, as well as supporting intelligent recommendations to encourage mutual responsiveness. F2: Users wanted actions to be grounded in personal narratives Participants felt that actions alone risked being too generic or ambiguous, and that pairing them with short narratives could add clarity and intimacy. As P3 noted, ”If I just send a pat on the shoulder, it feels flat. But if I can add a line like ’just like that day after your exam,’ it becomes ours.” Similarly, P1 remarked, ”Sometimes an action might be misread—like is this playful or serious? A few words can make the meaning clear.” Others emphasized the potential of narratives to embed personal identity and relationship specific themes. For instance, P2 explained, ”I’d love to add our inside jokes or little phrases we always use—it makes the chat feel more like us.” Overall, participants saw the combination of action and micronarrative as a way to transform fleeting gestures into more meaningful and accumulative exchanges. F3: Users desired flexibility to edit and control generated narratives Because relational communication is highly personal, participants expressed a strong desire for flexibility in shaping both actions and narratives. They wanted the ability to edit, refine, or override AI generated text, as well as to adjust persona cues that might influence how a message is framed. As P2 explained, ”Sometimes the AI sounds too formal—I just want to tweak it a little so it feels like my own voice.” Similarly, P4 remarked, ”I don’t want to be forced into one option. It’s important I can decline a suggestion or swap it for something that feels right for us.” These preferences highlight that users want AI to augment their expressiveness, but not to replace their agency. F4: Users expressed mixed preferences for ephemeral vs. persistent interactions Participants expressed a strong desire for flexibility in how expressive interactions are shared and retained. They wanted the option to send an action without any accompanying narrative, treating it as a lightweight and ephemeral gesture. At the same time, they emphasized the value of being able to choose whether an interaction should vanish quickly or persist as a lasting record. As P2 explained, ”Sometimes I just want it to disappear like a quick reaction, but other times I want it to stay so we can come back to it.” When interactions were made persistent, participants hoped they could also revisit the experience by replaying the associated action. P4 noted, ”If it’s saved, I’d like to click on it and see the moment again, not just read the text.” These preferences highlight users’ desire to balance spontaneity with continuity, allowing them to decide when an exchange should be transient and when it should become part of their shared history. 3.5. Design Goals Based on the findings from our formative study (F1–F4) and insights from prior literature on expressive messaging, social presence, and relational communication (Hassenzahl et al., 2012; Cassell, 2001; Birnholtz et al., 2012) we distilled the following design goals to guide the development of PuppetChat. Each design goal directly addresses one of the key findings. DG1. Support reciprocal action exchanges through a recommendation system. The system should provide interactive actions that invite reciprocal responses, supported by a recommendation mechanism. Based on a scoring process, the system should suggest four candidate actions tailored to the conversational context (e.g., recommending ”catch-heart” after ”throw-heart”), enabling playful turn taking and mutual responsiveness in dyadic chat. (F1) DG2. Generate micronarratives grounded in users’ personal stories. The system should create short, customizable narrative cues that draw on users’ personal story inputs and the selected action, allowing each message to carry contextually rich, relationship specific meaning. (F2, F3) DG3. Ensure editability and user control over narratives. The system should allow users to edit, override, or reject AI generated narratives and recommendations, ensuring personalization while preserving user agency. (F3) DG4. Allow flexible persistence and replay of interactions. The system should give users control over whether interactions remain ephemeral or are retained as revisitable records. When retained, interactions should also support an action replay function, allowing users to re-experience the associated gesture. This flexibility balances spontaneity with the creation of lasting and memorable exchanges. (F4) 4. PuppetChat System To address the 4 design goals identified in our formative study, we designed and implemented PuppetChat, a real time one to one conversational platform that augments instant messaging with reciprocal action exchanges and personalized micronarratives. PuppetChat augments instant messaging with reciprocal action exchanges paired with customizable micronarratives—short textual annotations that describe and contextualize each action in a personal way. One action–micronarrative pair provokes a reply, which in turn prompts another, creating a continuous chain of exchanges rather than isolated reactions. The primary user interface of PuppetChat, as depicted in Figure 2, serves as the central workspace for composing, sending, and responding to action–narrative messages. As shown in Figure 2, the main interface of PuppetChat consists of two primary components: the contact panel (Fig.2 A) and the conversation panel (Fig.2 B). The contact panel (Fig.2 A) enables users to add or search for contacts and manage login or presence status; within it, the relationship management module (Fig.2 A1) allows users to assign different icons to organize and differentiate among friends, family, partners, or varying degrees of closeness. The conversation panel (Fig.2 B) integrates multiple functions for dyadic interaction: the personal narrative input (Fig.2 B1) allows users to enter individualized cues and persona settings, which guide AI in generating and refining micronarratives; the interactive puppet area (Fig.2 B2) visualizes actions, where puppets remain in a resting pose when inactive and animate once an action is triggered, with the partner’s puppet displayed on the left and the user’s on the right; and the action recommendation button (Fig.2 B3) provides suggested actions based on conversational context, supporting reciprocal exchanges. Figure 2. The main interface of PuppetChat. (A) The contact panel supports user management. (A1) Relationship management through personalized icons. (B) The conversation panel for dyadic exchange. (B1) Personal narrative input for guiding and editing AI generated micronarratives with persona cues. (B2) The interactive puppet area, where puppets animate actions or remain idle in a resting pose (left: partner; right: self). (B3) Action recommendation button that suggests reciprocal actions based on conversational context. The main interface of PuppetChat. (A) The contact panel supports user management. (A1) Relationship management through personalized icons. (B) The conversation panel for dyadic exchange. (B1) Personal narrative input for guiding and editing AI-generated micronarratives with persona cues. (B2) The interactive puppet area, where puppets animate actions or remain idle in a resting pose (left: partner; right: self). (B3) Action recommendation button that suggests reciprocal actions based on conversational context. In this section, we describe the system design of PuppetChat in four parts. We begin by discussing how the action library was curated and annotated then outline key features of PuppetChat. Next, we conclude with the technical implementation that enables real time multimodal interaction. 4.1. System Design This section outlines the design rationale and implementation choices behind PuppetChat’s avatar and action system. 4.1.1. Avatar Design The finalized prototype uses a human-like girl avatar as its primary expressive character. This choice was informed by three considerations. (1) Our formative study showed that participants, regardless of gender, responded more positively to human-like characters than to robot or abstract avatars, and none of the male participants expressed discomfort with the girl avatar. (2) Prior research on 2D expressive assets, including stickers, emojis, and animated GIFs, indicates that user adoption is primarily driven by emotional clarity, stylistic expressiveness, and perceived enjoyment, rather than the specific gender of the character (Sawmong, 2022). In text based communication, these assets function as paralinguistic cues or pragmatic markers to modify tone, rather than as embodied self representations or strict identity claims (Tang and Hew, 2019; Aumüller et al., 2024). Consequently, the imperative for gender matching, while critical in embodied VR contexts to maintain presence, is significantly diminished in messaging systems. Evidence suggests that in these 2D environments, users prioritize functional utility and the ability to convey abstract concepts over demographic alignment, often preferring neutral or abstract representations when given the choice. (3) A further practical factor was technical feasibility. The system’s action library includes 42 expressive actions that require precise motion alignment, timing, and consistency. Generating and maintaining separate animation sets for multiple avatars would significantly increase production workload, model size, and integration complexity. Given our focus on evaluating core interaction mechanisms rather than character customization, we opted to standardize on a single human-like avatar to ensure consistency across actions and reduce implementation overhead. This decision mirrors common practice in animated sticker systems, which typically adopt a single stylized character for production efficiency. For these reasons, we adopted a single stylized human-like character, while acknowledging that offering additional avatar options is a valuable direction for future work. 4.1.2. Designing the Action Set The action library was constructed through a multi stage curation and annotation process. We began by collecting 215 animated GIFs featuring clearly identifiable protagonists and distinct actions, sourced from popular meme repositories as well as two dedicated GIF sharing websites111https://giphy.com/222https://tenor.com/. Each action was first categorized along two orthogonal dimensions: emotional valence (neutral, positive, negative) and interaction role (self oriented vs. responsive). This structure enabled the definition of potential ReactionCandidate pairings (e.g., ”throw-heart” can be paired with ”carry-heart” as a complementary response, while ”cry” can be paired with ”wipe others’ tears”). To ground these categories in everyday practice, we further collected over 20 conversation logs comprising nearly 100 instances of animated GIFs, emojis, and emoticons. We performed an inductive qualitative analysis on these instances, where two researchers independently coded the communicative intent of each artifact based on the surrounding textual discourse. From this process, we curated and refined a library of 42 representative actions, balancing expressive range with practical usability. Each action in our library is represented as a structured object with attributes such as name, description, keywords, emotion, interaction role, embedding array, and reactionCandidates; details are summarized in Table 1. Table 1. Attributes of each action in the library. This table lists the attributes defined for each action in our system’s library. Each action is represented as a structured object that includes: the Name, serving as the label of the action; the Description, a concise textual summary of the action’s content and expressive intent; Keywords, terms that capture typical triggers or contexts; the associated Emotion category; the Interaction role, situating the action within dyadic interaction; an Embedding array, a high-dimensional vector that enables semantic fuzzy matching; and ReactionCandidates, which link to complementary or reciprocal actions. Attribute Description Name The label of the action. Description A concise textual summary of the action’s content and expressive intent. Keywords A small set of terms that capture typical triggers or contexts. Emotion The associated affective category. Interaction role The situational function of the action within dyadic interaction. Embedding array A high dimensional vector enabling semantic fuzzy matching. ReactionCandidates Links to complementary or reciprocal actions. 4.2. Key Features In what follows, we detail PuppetChat’s key features, proceeding step by step along the workflow in Figure 3. Figure 3. Workflow of PuppetChat. (A) Everyday chat view. Pressing the Actions button (Fig. 2 B3) reveals (A1) a recommendation strip with four context aware actions sampled from a 42-item action library (A2), and clicking any action in (A1) advances to (B). (B) Composition view showing (B1) a visual preview of the selected action and an automatically generated micronarrative; tapping (B2) opens the customization panel (C). In (C), users refine tags via (C1) likes and dislikes, (C2) habits, (C3) social interaction style, and (C4) emotion, and may edit (C5) My Story. Clicking ”Use Tag to Regenerate” (C6) advances to regenerate the micronarrative based on the selected tags and enables sending. (A) Everyday chat view. Pressing the Actions button reveals (A1) a recommendation strip with four context-aware actions sampled from a 42-item action library (A2). (B) Composition view showing (B1) a visual preview of the selected action and an automatically generated micronarrative; tapping (B2) opens the customization panel (C). In (C), users refine tags via (C1) likes and dislikes, (C2) habits, (C3) social interaction style, and (C4) emotion, and may edit (C5) My Story. Clicking ”Use Tag to Regenerate” (C6) advances to regenerate the micronarrative based on the selected tags and enables sending. 4.2.1. Contextual Action Recommendation (scoring based) In our formative study, we found that participants often wanted to trigger puppet actions in two distinct ways: with input, when they typed either explicit commands or affective utterances, and without input, when they simply wished to react in the flow of conversation without sending additional text. Designing for both modes ensures that action recommendations remain responsive across different conversational circumstances. After the user clicks button B3 in Fig. 2, the system fuses text understanding and conversational context via a lightweight scoring scheme. For each candidate action a in the action library A (42 items), it computes a global score s(a∣c)=wtextstext(a∣c)+wctxsctx(a∣c)+wprefPreference(a,c)+ε(a),s(a c)=w_text\,s_text(a c)+w_ctx\,s_ctx(a c)+w_pref\,Preference(a,c)+ (a), where c is the current context, stext(a∣c)s_text(a c) is the text based score, sctx(a∣c)s_ctx(a c) is the context based score, Preference(a,c)∈[−1,1]Preference(a,c)∈[-1,1] is a user specific preference term, and ε(a) (a) is a small exploration noise term. In our implementation, we use fixed weights wtext=1.0w_text=1.0, wctx=1.0w_ctx=1.0, and wpref=0.5w_pref=0.5. When the user provides text they wish to send or express, the system performs a three layer interpretation to obtain a text based score stext(a∣c)s_text(a c) for each candidate action a. It first uses a language model to extract keywords and match them to action tags (Fig. 4 a), while performing negation detection to distinguish opposite-polarity statements such as ”I love you” (keyword/negation term +3+3; positive valence term +2+2, yielding stext≈+5s_text≈+5 before embedding adjustment) versus ”I don’t love you” (keyword/negation term −3-3 under negation; negative valence term +2+2, yielding stext≈−1s_text≈-1). It then performs emotional alignment (Fig. 4 b): the model determines the emotional valence of the input (positive, negative, or neutral) and matches it to the action’s emotion labels, contributing +2+2 for positive or negative alignment and +1+1 for neutral. Finally, when metaphorical or indirect expressions (e.g., ”I will give you an apple”) cause the first two steps to default to neutrality, the system computes the cosine similarity between embeddings of the input and the action descriptions/keywords to enable fuzzy semantic matching (Fig. 4 c), adding a small adjustment to stext(a∣c)s_text(a c). Users can trigger recommendations in two ways: they may enter text and then press the Actions button (Fig. 2 B3), or press B3 directly without entering text. In both cases, the contextual channel sctx(a∣c)s_ctx(a c) is evaluated in parallel to ensure responsiveness in reactive situations. As shown in Fig. 5 a, sctx(a∣c)s_ctx(a c) prioritizes complementary reactions by querying the action library’s ReactionCandidates with the partner’s most recent action (e.g., throw heart → catch/carry heart, hit with object → agony, cry → wipe other’s face); matching candidates receive a +5 bonus. Separately, Fig. 5 b applies an interaction role bias: each action is labeled as either self-oriented or responsive, and—conditioned on the ongoing conversational state—those marked responsive receive a +1 bonus to encourage reciprocal exchanges. A user specific preference term Preference(a,c)∈[−1,1]Preference(a,c)∈[-1,1] slightly promotes actions the user has frequently selected and demotes those they have repeatedly ignored or hidden. Together with the text based score stext(a∣c)s_text(a c), these contextual and preference contributions are combined via the scoring function above to yield the final ranking score s(a∣c)s(a c) used to select the top four recommendations. Figure 4. Text to action interpretation. (a) Keyword extraction with negation handling aligns user text with action tags; (b) emotional alignment maps input valence (positive, negative, or neutral) to the action’s emotion labels; (c) embedding-based fuzzy semantic matching covers indirect or metaphorical expressions. Text-to-action interpretation. (a) Keyword extraction with negation handling aligns user text with action tags; (b) emotional alignment maps input valence (positive, negative, or neutral) to the action’s emotion labels; (c) embedding- based fuzzy semantic matching covers indirect or metaphorical expressions. Figure 5. Contextual evaluation. (a) Reaction candidate matching: the partner’s most recent action queries the action library’s ReactionCandidates to prioritize complementary responses; (b) interaction role bias: actions labeled responsive (vs. self oriented) are promoted—conditioned on the ongoing conversational state—to encourage reciprocal exchanges. Contextual evaluation. (a) Reaction-candidate matching: the partner’s most recent action queries the action library’s ReactionCandidates to prioritize complementary responses; (b) interaction-role bias: actions labeled responsive are promoted—conditioned on the ongoing conversational state—to encourage reciprocal exchanges. 4.2.2. Micronarrative Generation and Adaptive Regeneration While actions enrich expressivity, our formative study showed that users sometimes sought added context or humor to make their actions feel more personal. Prior work on micronarratives suggests that short, situated captions can deepen social presence and support lightweight identity work in digital exchanges (Ahmmed et al., 2024; Oh et al., 2018). Building on this insight, PuppetChat introduces micronarratives: brief, context aware captions that accompany actions and are generated by drawing on three aspects: (1) the user’s personal story, a short self description provided voluntarily (entered in Fig. 2 B1); (2) the selected action, which serves as the expressive trigger; and (3) the surrounding conversational context, which situates the action in dialogue. When no personal narrative is available, the system defaults to a neutral baseline to preserve coherence. In the workflow, once a user selects any of the four recommendations in Fig. 3 A1, the composition view in Fig. 3 B displays the AI-generated micronarrative aligned with that action. Tapping B2 opens the customization panel in Fig. 3 C, where C1 (likes and dislikes), C2 (habits), C3 (social interaction style), and C4 (emotion) provide lightweight controls for refinement. Conditioned on the personal narrative in Fig. 2 B1, the system proposes five candidate tags within each category. Users may select from these or add their own. If inspired by the ongoing chat or by chosen tags, users can revise their personal narrative via C5, after which pressing C6 produces a regenerated micronarrative. Within this process, keywords extracted from the personal narrative act as persona anchors that help maintain a consistent voice across turns while staying sensitive to the current action and context. For example, given the tag ”cat” and the action wipe-tears, the regenerated caption might read: ”I’l be your gentle cat paw wiping away those tears.” As another example, with the tag ”marathon” and the action high-five, the caption could be: ”A high-five at mile 26, we made it.” These adjustments make captions feel individualized without imposing heavy authoring overhead. This functionality operationalizes DG2 (integrate actions with personalized micronarratives) and DG3 (ensure editability and user control over narratives): micronarratives are generated automatically for immediacy yet remain easily rewritable through tags and personal narrative edits. By aligning captions with user provided anchors and conversational context, it also complements DG1 by supporting responsive, meaningful exchanges in dyadic chat. 4.2.3. Reciprocal Action Exchange From the workflow controls (Fig. 3 B), the user can either choose ”Action Only” or proceed to ”Send.” Selecting ”Action Only” dispatches the chosen action without its micronarrative. The event appears as the ephemeral status view in Fig. 6 (a) and leaves no micronarrative trace in the log, aligning with DG4 (ephemeral, low commitment expression). Choosing ”Send” transmits both the action and the edited micronarrative. The conversation history preserves a persistent textual record as in Fig. 6 (b). Tapping this record replays the original action rendering. When the partner responds, the status transitions to a dyadic exchange (Fig. 6 (c)), in which both puppets are co-present and their actions are rendered as an interaction. The exchange is appended to the thread, supporting turn taking and continuity in the chat history. Figure 6. Status views for action delivery and exchange. (a) Action Only: the selected action is dispatched without a micronarrative. (b) Send: the action and the edited micronarrative are delivered together. (c) Partner response: both puppets are co-present and their actions are rendered as an interaction. Status views for action delivery and exchange. (a) Action Only: the selected action is dispatched without a micronarrative. (b) Send: the action and the edited micronarrative are delivered together. (c) Partner response: both puppets are co-present and their actions are rendered as an interaction. 4.3. Implementation 4.3.1. Frontend Implementation The frontend is built with Vue 3333https://vuejs.org/ (Composition API) and styled using Tailwind CSS444https://tailwindcss.com/ for a responsive, utility first interface. It comprises three main components: an action palette, a message input box with micronarrative previews, and a real time activity feed. Client–server communication is maintained via WebSockets555https://socket.io/ (Socket.IO), ensuring actions and recommendations are broadcast to conversation participants with end-to-end latency of about 200 ms. 4.3.2. Backend and AI Components The backend uses Node.js666https://nodejs.org/zh-cn and Express777https://expressjs.com/, with REST endpoints for user/session management, CRUD over the action library, and micronarrative generation. Real time events (e.g., puppet-action, emn-update) are broadcast to dyad specific namespaces. Data are stored in MongoDB across collections for actions, conversations, users, logs, and messages. Secondary indexes and attached embeddings (via Azure NLP API888https://azure.microsoft.com/) enable semantic retrieval, while Azure OpenAI supports emotion/role inference and micronarrative generation. Recommendations combine heuristics, embedding similarity, and LLM annotations. Results are logged and broadcast, with options for action only or action and micronarrative persistence. 5. User Study We conducted a 10-day field study with 11 dyads (22 participants) in close relationships to assess how PuppetChat supports reciprocal action exchanges, personalized micronarratives, and everyday communication999The study protocol was reviewed and approved by the IRB at the lead author’s institution, where the study was conducted.. 5.1. Participants We recruited 11 dyads (22 participants; 12 women, 10 men; ages 21–28, M = 24.45, SD = 1.95) in close relationships. Dyads self identified as best friends (6) or couples (5). Participants spanned student and professional roles across HR/business, software engineering, data analytics, banking, nursing, and marketing. To reflect everyday habits across ecosystems, participants reported a heterogeneous mix of messaging applications (e.g., WhatsApp, Instagram, WeChat, Telegram, Discord, LINE, iMessage). We refer to dyads as P1–P11 and to members as A/B; detailed profiles appear in Table 6. 5.2. Procedure Before deployment, each participant completed a baseline questionnaire capturing demographics, commonly used messaging apps, typical communication frequency with a close friend or partner, typical conversation duration, and expected response latency. Participants then used PuppetChat in situ with their existing partner or close friend for 10 consecutive days. We provided 10 conversation topics; on Days 1–6, participants selected one topic per day and chatted for at least 15 minutes, while on Days 7–10 they could either continue with listed topics or chat freely without constraints. During deployment, the system logged 5 fine grained interaction events. In parallel, participants completed a brief two-item micro survey once per day about that day’s chatting experience and technical issues. On Day 10, a summative questionnaire assessed Co-presence and Affective Intimacy (DG1, DG2), Reciprocal Interaction (DG1), Personalized Micronarratives (DG2), and Agency and Persistence (DG3, DG4), together with the System Usability Scale (SUS) (Brooke, 1996; Bangor et al., 2009) and NASA–TLX (Hart, 2006) workload. Following the deployment, each dyad participated in a 40–60 min semi-structured interview to probe experiences. 5.3. Measures and Data Analysis 5.3.1. Quantitative Analysis and Instrument Design Backend telemetry was aggregated into daily and dyad level usage metrics, including action frequency, reciprocal exchange patterns, micronarrative edits, and persistence choices. These data were analyzed using descriptive statistics to capture engagement trends over the 10-day deployment. Standardized usability and workload were assessed using SUS and NASA–TLX scores, computed following standard guidelines. To capture the specific relational and expressive affordances of the system, we developed a custom 16-item instrument. The design of this instrument followed a deductive scale development approach. We first drew on Social Presence Theory to operationalize co-presence and affective intimacy (Kreijns et al., 2022; Cui et al., 2013). For reciprocal interaction, we adapted ideas from turn taking and entrainment in interpersonal coordination (Walther, 1992, 1996). Items for personalized micronarratives were informed by research on narrative self disclosure and story based communication in close relationships (Bizzocchi et al., 2014). Finally, our measures of agency and persistence were guided by frameworks in AI Mediated Communication that foreground the balance between algorithmic assistance and user control (Walther, 1996). Each item was then refined to fit the specific context of PuppetChat and mapped to one of four dimensions: Co-presence and Affective Intimacy (DG1, DG2), Reciprocal Interaction (DG1), Personalized Micronarratives (DG2), and Agency and Persistence (DG3, DG4). The Basic scale consisted of sixteen items organized into four subdimensions. Co-presence and Affective Intimacy captured participants’ sense of mutual presence and closeness, including perceived partner presence (Presence), willingness to disclose genuine feelings (Disclosure), the emergence of shared rituals (Rituals), and a warmer relational atmosphere (Warmth). Reciprocal Interaction assessed the mutuality of expressive exchanges, including paired or complementary actions (Pairs), understanding of action intentions (Understanding), appropriateness of recommended reciprocal actions (Fit), and the rhythmic quality of matched interactions (Rhythm). Personalized Micronarratives reflected how system generated textual elements enriched continuity, including natural links to past experiences (Connection), the construction of coherent stories (Story), the possibility of reusing narratives (Reuse), and a reverse coded item indicating when narratives felt empty or templated (Empty). Finally, Agency and Persistence captured users’ sense of control and temporal durability, including the ease of editing system outputs (Editing), feelings of autonomy (Control), the ability to preserve meaningful interactions (Memory), and a reverse coded item describing ephemeral disappearance of expressions (Ephemeral). The final questionnaire items were derived from these same theoretical frameworks, and the complete set of questions is presented in Table 5. 5.3.2. Qualitative Analysis and Integration For the qualitative component, we analyzed interviews and open-ended survey responses using a reflexive thematic analysis approach (Braun and Clarke, 2019). Authors independently reviewed all transcripts to generate initial codes related to action exchange, micronarrative use, editing behaviors, friction points, and expectations around persistence. Through iterative discussion, the authors refined these codes, resolving discrepancies and consolidating them into higher level themes. 6. Study Results We report our study results by combining quantitative snapshots of system use with findings from post study interviews. 6.1. PuppetChat Usage (N=22) Participants’ usage logs indicate sustained engagement in composing and refining ”action + micronarrative” messages. On average, participants pressed the Actions button (Fig. 2 B3) 18.0 times each (M = 18.0; median = 17.5; IQR = 13–22.5; range = 8–34) and ultimately sent 14.5 actions (M = 14.5; median = 14; IQR = 10.5–18; range = 6–28), yielding an ∼ 80% trigger→ conversion. Micronarratives accompanied (Fig. 6 (c)) 8.2 messages on average (M = 8.2; median = 8; IQR = 5–11; range = 2–16), i.e., ∼ 57% of sent actions. Effort was also evident in editing and ”peek behind the scenes” (Fig. 3 B2): participants carried out 13.2 edit operations on average (M = 13.2; median = 12.5; IQR = 9–16.5; range = 5–27), and partners opened 9.7 previews on average (M = 9.7; median = 9.5; IQR = 7–12; range = 3–18). Usage varied widely across individuals—for example, U13 was the heaviest user (28 actions, 16 micronarratives), whereas U20 used the system sparingly (6 actions, 2 micronarratives). Table 2. PuppetChat usage by participant (original logs). PuppetChat usage by participant (original logs). Participant Total Action Micronarr. Edit ops Preview opens U01 26 20 12 19 15 U02 18 14 9 13 11 U03 22 17 10 21 12 U04 15 12 5 9 10 U05 19 16 11 18 14 U06 13 10 6 12 8 U07 21 17 9 10 9 U08 12 9 4 7 6 U09 24 21 13 17 10 U10 16 13 7 15 9 U11 11 8 3 6 5 U12 20 16 8 14 9 U13 34 28 16 27 18 U14 14 11 6 10 7 U15 17 14 8 12 11 U16 23 19 12 16 13 U17 10 8 5 9 4 U18 9 7 4 6 3 U19 27 22 14 20 16 U20 8 6 2 5 4 U21 25 19 11 15 12 U22 13 11 5 9 7 6.2. Quantitative Snapshot from the Basic Scale This section reports our results about how participants used editable micronarratives to express subtle emotions and personalize shared meaning. 6.2.1. Reliability and Descriptive Statistics We first assessed the internal consistency and descriptive statistics of the four subscales (Table 3) and the overall 16-item instrument(Fig. 7). Cronbach’s α values indicated acceptable to excellent reliability across all subscales, with coefficients ranging from .78 to .86, and very high reliability for the full scale (α=.93α=.93). This suggests that the items within each construct measured coherent underlying dimensions. Descriptive analyses showed that participants evaluated the system positively across all dimensions, with mean ratings ranging from 5.1 to 5.4 on a 7-point Likert scale. The overall scale mean was 5.26 (SD=1.00SD=1.00), reflecting generally favorable experiences. Floor effects (scores ≤2≤ 2) were negligible, while ceiling effects (scores ≥6≥ 6) were observed in approximately 10–20% of responses depending on the subscale. These results suggest that while responses were positively skewed, there was sufficient variability to differentiate between participants’ experiences. Table 3. Reliability and descriptive statistics for each subscale and the overall instrument. Reliability and descriptive statistics for each subscale and the overall instrument. Scale α M SDSD Co-presence & Affective Intimacy .78 5.3 0.9 Reciprocal Interaction .82 5.1 1.0 Personalized Micronarratives .86 5.2 1.1 Agency & Persistence .83 5.4 0.9 Total (16 items) .93 5.26 1.0 Figure 7. Distribution of responses for the 16 items of the Basic scale (N=22). Each boxplot shows the median and interquartile range of item ratings on a 1–7 Likert scale. reverse coded items are labeled with “(R)”; for all other items, higher scores indicate stronger endorsement. Distribution of responses for the 16 items of the Basic scale (N=22). Each boxplot shows the median and interquartile range of item ratings on a 1–7 Likert scale. Reverse-coded items are labeled with “(R)”; for all other items, higher scores indicate stronger endorsement. Figure 7 presents a box plot illustrating these results. Overall, item level means indicated that participants most strongly endorsed the persistence related item of resisting ephemerality (Ephemeral, M=6.27M=6.27), suggesting a clear preference for maintaining interaction records rather than letting them vanish. By contrast, the lowest endorsement was observed for the micronarrative item linking to past experiences (Connection, M=4.27M=4.27), highlighting limited resonance between system generated narratives and participants’ personal histories. Other highly rated items included Disclosure, Understanding, and Memory (M>5.6M>5.6), underscoring that PuppetChat effectively fostered openness, reciprocal understanding, and the desire to preserve meaningful interactions. These results align with the design goals, showing particularly strong support for DG3–DG4 (Agency and Persistence), while also pointing to opportunities for refining DG2 (Personalized Micronarratives). 6.2.2. System Usability and Workload To situate our system within established benchmarks, we administered the System Usability Scale (SUS; 0–100) and the NASA–TLX workload inventory. For the latter, we adopted the raw workload index (RTLX), calculated as the mean of the six oriented NASA–TLX subscales (0–100, with the Performance scale reversed). Table 4 summarizes the results across all participants (N=22N=22). Table 4. SUS and NASA–TLX overview. SUS and NASA–TLX overview. Measure N Mean SD 95% CI Low 95% CI High SUS 22 86.59 17.77 68.71 84.47 RTLX 22 31.72 6.98 28.63 34.81 The mean SUS score was 86.5986.59 (SD=17.77SD=17.77), with a 95% confidence interval of [68.71,84.47][68.71,84.47]. This value substantially exceeds the conventional benchmark of 68, and falls into the ”excellent usability” range commonly reported for SUS scores above 80. In contrast, the mean RTLX workload score (Fig. 8) was relatively low (M=31.72M=31.72, SD=6.98SD=6.98, 95% CI [28.63,34.81][28.63,34.81]), indicating that interaction with the system required only modest cognitive and physical effort. Figure 8. Box plots of the six NASA–TLX dimensions. Box plots of the six NASA–TLX dimensions. Taken together, these findings suggest that participants experienced PuppetChat as both usable and lightweight to operate. The high SUS scores align with the positive impressions reported in the qualitative feedback, while the low NASA–TLX scores confirm that the system imposed little additional workload beyond the conversational task itself. 6.3. Post-Study Interviews Our post-study interviews with participants further reveal how they relied on PuppetChat to maintain continuity over time and reflect on evolving shared histories. 6.3.1. Cultivating Social Presence through Reciprocal Co-performance Establishing social presence hinges on a focused, distraction free communication space and a strong expectation of prompt responses. Compared with general purpose messengers, PuppetChat’s single purpose environment and the ”sleeping puppets” cue reduce distraction and heighten joint attention. Participants also reported a strong expectancy that messages would be seen and answered promptly, which stabilized the tempo of interaction and strengthened the felt sense of being co-located. Such as P1-A said ”I really like the two little figures who are always asleep, they make my entry feel meaningful: I can wake them up, which gives me a more focused feeling. Opening PuppyChat feels like stepping into a room that’s just ours, no pop-ups, no other chats. ”(P1-A). Another participant reported the same experience, felt a strong sense of confidence and expectation that her partner would quickly notice and respond to her actions. Such as P2-B said ”Because I know that once my action is sent, the other person can respond, and will get many recommended ways to respond. I expect and am confident they’l reply. ” (P2-B) Fostering anticipated reciprocity, action response grammar transforms individual expressions into shared, jointly completed performances. Participants understood the action response grammar as an invitation closure sequence: sending an action initiates a jointly completed performance rather than a solitary broadcast. System supported complementary or mirrored replies make being answered feel probable at the moment of sending, shifting the experience to ”we did something together.” As P6-A reflected ”Lately I’ve been exhausted, but I sent a knees-hug action [action: knees-hug] in this app. Getting a hug-back feels like we did something together, not just pasted a sticker. When I sent it, it felt less like expressing my emotion and more like showing my state, and I knew the other person would be there with me.” (P6-A) Reanimating shared routines and role complementarity, PuppetChat’s co-performance features enable the represencing of familiar life scripts, thereby recalling the embodied sense of us even across distances. Co-performance is not only emotional, it reanimates shared routines and role complementarity, recalling the embodied ”us” despite distance. We found that PuppetChat’s co-performance features enabled the represencing of familiar roles and shared routines, creating a strong sense of her partner’s presence and recalling past embodied experiences despite physical distance. P10-A said ”Compared to other instant messengers, I feel their presence more strongly here. We hadn’t seen each other in person for a long time. I said I had taken many selfies, so I sent a ‘fan-self’ action. She replied with ‘take photo,’ and I felt like she was still by my side, just like before, my dedicated photographer.” (P10-A) Cultivating social presence in intimate chat involves establishing a focused and low distraction environment with reliable expectations of reciprocal replies. Upon this foundation, action response turns convert individual expressions into reciprocal, closed sequences of co-performance. This shift is strongly validated by high mean scores on Co-performance & Rhythm (M=5.34) and Presence & Warmth (M=5.24) scales, which also show a strong positive correlation with system usability. 6.3.2. Expression and Personalization Editable micronarratives allowed participants to articulate subtle feelings that might otherwise remain unspoken, providing a lightweight way to preserve face while reducing ambiguity. We found that participants used this feature to disclose dissatisfaction or delicate emotions indirectly, articulating what they might hesitate to say outright. P5-A wanted to hint that he was a bit annoyed after his best friend forgot to update him about their weekend plan. Rather than saying this outright, he chose the “agony” action and edited its micronarrative from the dramatic default to a teasing, understated line: “someone went missing today…” This combination let him express himself in a joking and concise way that fit their usual banter. As he explained, ”I’m not someone who speaks very directly, so I edit the short description to express that I’m a little unhappy.” (P5-A) Micronarratives also enabled participants to cultivate shared symbolic codes and playful shorthand, reinforcing intimacy and a sense of we-ness. By pairing actions with personalized text, users generated dyadic ”inside jokes” that recalled shared identity and history. We found that such practices generated micro cultures of communication, where reciprocal riffs around everyday dislikes or preferences strengthened relational closeness beyond what emojis or text alone could achieve. As one participant noted, ”We both really hate cilantro, so once I sent a vomiting action, and after she replied with the same action, both of our short descriptions were about cilantro. I was like, wow, this is our tacit understanding, no wonder we’re such good friends.” (P9-B) Participants diverged in how much effort they invested in editing micronarratives, revealing two contrasting orientations toward use. For some, micronarratives resembled avatars or profile bios, serving as lightweight surfaces to display individuality and highlight shared affinities. These participants embraced personalization as a way to signal both their personality and their relational commonalities. As P4-A put it, ”I love avatars and profile bios, this micronarrative feels the same. It helps me achieve personalization. On one hand it connects us through common points, but it also lets me show my personality.” (P4-A) In contrast, others were less inclined to modify text, preferring brevity or shifting emphasis toward the partner rather than themselves. For these participants, too much self related content felt inappropriate or self centered, and they valued contextually relevant prompts that expressed care and attention to the other. As one participant explained, ”I don’t really like too much content about my personal information—maybe because I’m the one who fawns over the other person, I often want to express my attention to them. Showing too much of myself can feel self centered. It would be better if it connected more with the conversation context.” (P10-B) Taken together, these accounts show how PuppetChat supported both personalization oriented identity work and other directed, low effort expression, accommodating different relational and communicative preferences. This resonates with our quantitative results: higher scores on Editing, Control, and Memory (DG3–DG4) confirm strong agency and persistence, while comparatively lower ratings on Connection (DG2) highlight the design opportunity to strengthen ties between generated text and users’ lived experiences. 6.3.3. Continuity and Reflection Beyond immediate expression, participants emphasized the importance of continuity and reflection in sustaining meaningful interaction. They viewed PuppetChat not only as a channel for moment-to-moment communication but also as a medium for capturing evolving moods, shared histories, and ongoing narrative arcs. Participants highlighted the need for lightweight tools to maintain continuity, both in tracking daily moods and in preserving a sense of shared history. Some wished for a ”save multiple versions” function for personal stories, treating them as day level state markers that could be easily adjusted without reauthoring from scratch. This versioning function was seen as both expressive, capturing daily shifts in mood, and ergonomic, reducing authoring burden. As one participant explained, ”I wish there were a save button in personal story so I could save my different stories for each day. I treat this description as a reflection of my daily state.” (P3-B). Others emphasized retrospective viewing, using past entries as a jointly owned archive that documents how the relationship evolves over time. As another participant described, ”I revisit past entries because I want to see how my partner and I were in different states at different times—it’s a way of preserving memories that belong to us.” (P6-A). Continuity also emerged in the form of narrative arcs, where actions and short descriptions clustered around focal themes over stretches of time. Participants described how prior moves seeded subsequent ones, producing a serialized flow that sustained momentum and deepened engagement with a shared topic. As one participant reflected, ”This continuous record feels like a series between us, and I realized our discussion felt like a serialized show. Seeing the descriptions in her replies gave me lots of ideas and made me want to respond.” (P5-A). Taken together, these accounts demonstrate how continuity and reflection broadened PuppetChat’s role from facilitating single turns of expression to sustaining longer term relational trajectories. The system fostered a sense of temporal depth and shared narrative, reinforcing both personal presence and collective identity over time. 7. Discussion In this section, we interpret our findings, examine how PuppetChat supports reciprocal and expressive communication, outline key design implications, and then reflect on limitations of our study and directions for future work. 7.1. From Terminal Feedback to Reciprocal Co-Performance A central finding of our work is that intimacy in mediated environments is structurally supported by invitational affordances, interface mechanisms that solicit and structure reciprocal responses rather than by mere visual richness. Prior literature on IM often highlights the prevalence of ”terminal feedback”, such as reaction emojis that signal acknowledgement but effectively close a thread (Woerner et al., 2007; Schwarz, 2011). In contrast, PuppetChat’s complementary action recommendation system (Fig. 5 (a)) offers a concrete instantiation of a grammar of interdependence: suggested responses are framed as follow-ups that fit the partner’s move rather than as isolated reactions. The quantitative results, particularly the high scores on Reciprocal Interaction (M=5.1M=5.1) and Co-presence (M=5.3M=5.3), are consistent with the interpretation that users perceived these exchanges not as isolated signals, but as coupled units of behavior. We interpret this shift as the emergence of asynchronous co-performance: a form of joint action in which temporally staggered, yet mutually responsive, turns accumulate into a shared performance. Qualitative accounts of the ”sleeping puppets” creating a ”focused room” suggest that the interface staged a dormant co-presence—a state of readiness that elevated the expectation of a reply even when partners were not strictly co-temporal. Unlike standard stickers, where the choice is vast and often arbitrary, the structured recommendations narrowed the design space to actions that ”fit” the partner’s previous move (e.g., distress → comfort). This constraint appeared to increase the perceived meaningfulness of the interaction: by reducing the cognitive load of selecting a response, the system allowed users to focus on the rhythm of the exchange. This reading aligns with theories of joint action (Podkosova et al., 2023), suggesting that in remote contexts, the feeling of ”being together” can arise less from strict simultaneity and more from the complementarity of contributions. The same invitational structure appears to underpin continuity. Rather than terminating an exchange, reciprocal chains of actions and replies keep threads alive across moments, while editable micronarratives let users quickly tune meaning without reauthoring from scratch. When users choose to persist action and micronarrative pairs, these micro sequences stitch across sessions into lightweight, revisitable traces. This pattern is consistent with prior work on how persistent traces and joint reminiscence scaffold shared memory and social presence (Roediger and Pyc, 2012; Ahmmed et al., 2024), showing that selected exchanges become a jointly owned archive that can be replayed to re-experience prior moods and rhythms and to pick up ongoing threads with a sense of being seen across time. Our deployment included both romantic partners and best friends, all of whom described their relationships as already close. We did not design or power the study to test systematic differences between relationship types, and we observed no robust quantitative divergences in how reciprocal actions or continuity features were taken up between close friends and romantic partners. Any descriptive differences (e.g., in thematic content or humour style) were minor. We therefore cautiously treat these mechanisms as applying to close dyads more generally, while acknowledging that relationship type, history, and stage of intimacy may modulate how strongly specific features (e.g., playful co-performance vs. reflective revisiting of traces) are used. A finer grained comparison of early stage versus established relationships, and of different relational forms, remains an important direction for future work rather than a claim of this study. 7.2. Micronarratives: Ambiguity and Specificity While embodied actions provided the vehicle for co-presence, micronarratives—short, AI suggested captions paired with those actions provided the semantic anchor. Our results indicate a complex relationship between AI generated text and user intent. The ”cilantro” example (Section 6.3.2), in which a generic template about food was appropriated into a highly specific inside joke, illustrates how dyads repurposed generic suggestions to construct private meaning. This supports the notion that intimacy tools should leave room for multiple interpretations (Baxter and Bullis, 1986), but adds a nuance: AI generation in this setting serves as a scaffold for appropriation rather than a final product. A potential concern is that AI generated micronarratives might feel impersonal or disconnected from the relationship, or that reliance on such suggestions could erode people’s ability to craft expressive conversations on their own. The relatively modest score on the Connection item, which captured how well the system linked to past experiences, reflects this limitation and suggests that generic templates alone may not fully convey relational history. At the same time, participants frequently engaged with the editing tools and reported high willingness to disclose, suggesting that they did not treat the initial AI output as a finished message. Instead, “imperfect” suggestions functioned as provocative resources that invited users to articulate and refine feelings they might otherwise leave unspoken. Here, the value of micronarratives lies less in the accuracy of zero shot generation and more in lowering the activation energy for nuanced, face saving self disclosure, making it easier to start and adjust a vulnerable message without feeling overly exposed. Although our 10 day deployment cannot address long term skill change or subtle shifts in authorship, the observed pattern points toward scaffolding rather than substitution: meaningful exchanges emerged from users’ active reshaping of generated text, indicating that partners continued to practice and elaborate their own relational language rather than outsourcing it to generative AI. Longer term, comparative studies are needed to examine how such scaffolds affect people’s expressive repertoires over time. 7.3. Design Implications The design and evaluation of PuppetChat provide several design implications for HCI and CMC systems seeking to support remote intimacy: • Embed embodied actions into messaging turn structure with complementary grammars. Rather than treating gestures, animations, or haptics as one-off reactions, interfaces can define simple complementary patterns (e.g., distress → comfort, offer → acceptance) that make a reply feel both easy and expected, turning individual expressions into co-performances. Our participants’ descriptions of ”sleeping puppets” and back-and-forth exchanges suggest that such patterns can support a felt sense of asynchronous togetherness. • Offer a lightweight, editable middle layer of expression. Short, regenerable captions paired with actions can bridge the gap between emojis and full text, especially when they support face saving specificity and inside jokes. This layer should be easily skippable for users who prefer minimalism, but powerful for those who wish to craft a distinctive persona and shared humour, as reflected in how dyads appropriated and edited micronarratives in our study. • Prioritize user agency in AI mediation. AI should propose rather than impose: recommendations must be editable, overridable, and tunable in intensity, with clear affordances for decline. Grounding generation in user supplied tags and stories respects privacy while still enabling personalization. Our observations of frequent editing and selective uptake highlight the importance of keeping users in control of how AI generated text enters their relational language. • Design persistence as narrative infrastructure. Giving users nuanced control over which interactions persist and how they can be replayed, grouped, or revisited can help transform momentary exchanges into a shared history. Simple mechanisms such as replay on click and optional story versioning can support reflection and joint reminiscence without heavy authoring, echoing our participants’ use of persisted action+micronarrative pairs as lightweight, revisitable traces. Although our deployment involved a mix of romantic partners and close friends, the underlying mechanisms proved meaningful across relationship types as discussed in section 7.1. Rather than being tied to a specific relational category, these implications point to generalizable interaction principles for supporting care, play, and a sense of co presence at a distance. 7.4. Limitations and Future Work While our field study deployment demonstrates the potential of PuppetChat in enhancing intimate communication among already close dyads, several limitations remain. First, the relatively small sample size of 22 participants, recruited primarily from a university community and consisting of long term friends and romantic partners, constrains the generalizability of our findings. We did not systematically compare different relationship forms or stages of intimacy, nor did we explicitly recruit neurodivergent participants or attend to intercultural variation in depth. Second, the deployment spanned only ten days, so we cannot speak to the long term impact of reciprocal actions and story driven micronarratives on relationship trajectories, nor can we assess whether people might come to over rely on micronarratives in ways that affect their in person communication. Finally, the study evaluated the integrated system as a whole without comparison conditions, making it difficult to identify which specific elements (e.g., ReactionCandidate, micronarrative generation, editing tools, persistence settings) contributed most strongly to the observed effects. Future work will therefore expand both scope and granularity. We plan to broaden the participant pool to include a more diverse range of partnerships (e.g., family members, intercultural and cross border relationships), and to examine how different relationship types and intimacy stages modulate the uptake and impact of co-performance and continuity mechanisms. Although our present study does not speak directly to these populations, we see particular promise in studying how action and micronarrative scaffolds can support neurodivergent individuals who may find it challenging to communicate or interpret tone and intent in text based messaging, while carefully attending to accessibility and agency. Cross cultural deployments will be important for understanding how norms of expressivity, face saving, and reciprocity shape the design space. Methodologically, we aim to conduct longitudinal studies to assess PuppetChat’s longer term influence on relational dynamics, and to perform ablation and comparison studies (e.g., disabling complementary recommendations, removing micronarratives, or varying persistence) to clarify which mechanisms, individually and in combination, are most impactful for future system design. 8. Conclusion In this paper, we present PuppetChat, a dyadic messaging prototype that enriches intimate communication by addressing the limits of one-off reactions. Users express with lightweight character actions coupled to editable, context aware micronarratives, while the system embeds structured reciprocity by surfacing complementary or mirrored responses at send time, turning solitary expressions into reciprocal sequences. Selective record and replay then supports continuity and shared memory. Our user study demonstrated that PuppetChat supports users in initiating richer conversations and adapting AI suggestions to their own relational tone, as well as maintaining a sustained sense of connection over time, resulting in more engaged and meaningful exchanges compared to everyday messaging practices. References (1) Adler et al. (2010) Rhodes Adler, Athanasia Vasiliadis, and Nina Bickell. 2010. The relationship between continuity and patient satisfaction: a systematic review. Family Practice 27, 2 (01 2010), 171–178. doi:10.1093/fampra/cmp099 Ahmmed et al. (2024) Asif Ahmmed, Erica Butts, Kimia Naeiji, Ladda Thiamwong, Julie Ancis, and Salam Daher. 2024. Aware Intelligent Virtual Agent’s Effect on Social Presence and Empathy of Caregivers. In Proceedings of the 24th ACM International Conference on Intelligent Virtual Agents (GLASGOW, United Kingdom) (IVA ’24). Association for Computing Machinery, New York, NY, USA, Article 27, 5 pages. doi:10.1145/3652988.3673951 Ale-Ebrahim et al. (2023) Benjamin Ale-Ebrahim, Tristan Gohring, Elizabeth Fetterolf, and Mary L. Gray. 2023. Pronouns in the Workplace: Developing Sociotechnical Systems for Digitally Mediated Gender Expression. Proc. ACM Hum.-Comput. Interact. 7, CSCW1, Article 83 (April 2023), 30 pages. doi:10.1145/3579516 Alsayahani (2023) Hatim Alsayahani. 2023. Overcoming Customisation Challenges in Information Dashboards. In Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization (Limassol, Cyprus) (UMAP ’23). Association for Computing Machinery, New York, NY, USA, 285–289. doi:10.1145/3565472.3595646 Ames et al. (2010) Morgan G. Ames, Janet Go, Joseph ’Jofish’ Kaye, and Mirjana Spasojevic. 2010. Making love in the network closet: the benefits and work of family videochat. In Proceedings of the 2010 ACM Conference on Computer Supported Cooperative Work (Savannah, Georgia, USA) (CSCW ’10). Association for Computing Machinery, New York, NY, USA, 145–154. doi:10.1145/1718918.1718946 Andalibi et al. (2017) Nazanin Andalibi, Frank Bentley, and Katie Quehl. 2017. Multi-Channel Topic-Based Mobile Messaging in Romantic Relationships. Proc. ACM Hum.-Comput. Interact. 1, CSCW, Article 20 (Dec. 2017), 18 pages. doi:10.1145/3134655 Andersen et al. (1979) Janis F Andersen, Peter A Andersen, and Arthur D Jensen. 1979. The measurement of nonverbal immediacy. Journal of applied communication research 7, 2 (1979), 153–180. Aumüller et al. (2024) Anna Aumüller, Andreas Winklbauer, Beatrice Schreibmaier, Bernad Batinic, and Martina Mara. 2024. Rethinking feminized service bots: user responses to abstract and gender-ambiguous chatbot avatars in a large-scale interaction study. Personal and Ubiquitous Computing 28, 6 (2024), 1021–1032. Bangor et al. (2009) Aaron Bangor, Philip Kortum, and James Miller. 2009. Determining what individual SUS scores mean: Adding an adjective rating scale. Journal of Usability Studies 4, 3 (2009), 114–123. Baxter and Bullis (1986) Leslie A Baxter and Connie Bullis. 1986. Turning points in developing romantic relationships. Human communication research 12, 4 (1986), 469–493. Bhandari and Bardzell (2008) Shruti Bhandari and Shaowen Bardzell. 2008. Bridging gaps: affective communication in long distance relationships. In CHI ’08 Extended Abstracts on Human Factors in Computing Systems (Florence, Italy) (CHI EA ’08). Association for Computing Machinery, New York, NY, USA, 2763–2768. doi:10.1145/1358628.1358758 Birnholtz et al. (2012) Jeremy Birnholtz, Jeff Hancock, Madeline Smith, and Lindsay Reynolds. 2012. Understanding unavailability in a world of constant connection. Interactions 19, 5 (2012), 32–35. Bizzocchi et al. (2014) Jim Bizzocchi, Michael Nixon, Steve DiPaola, and Natalie Funk. 2014. The Role of Micronarrative in the Design and Experience of Digital Games. https://doi.org/10.26503/dl.v2013i1.669. In Proceedings of DiGRA 2013 Conference. DiGRA, Tampere. Bødker and Grønbæk (2020) Susanne Bødker and Kaj Grønbæk. 2020. Design in action: From prototyping by demonstration to cooperative prototyping. In Design at work. CRC Press, 197–218. Braun and Clarke (2019) Virginia Braun and Victoria Clarke. 2019. Reflecting on reflexive thematic analysis. Qualitative research in sport, exercise and health 11, 4 (2019), 589–597. Brooke (1996) John Brooke. 1996. SUS: A quick and dirty usability scale. Technical Report TR 26.1. Digital Equipment Corporation, Reading, UK. Available in: Jordan, P.W., Thomas, B., Weerdmeester, B.A., & McClelland, A.L. (Eds.) (1996). Usability Evaluation in Industry. London: Taylor & Francis.. Bruns et al. (2021) Miguel Bruns, Stijn Ossevoort, and Marianne Graves Petersen. 2021. Expressivity in Interaction: a Framework for Design. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI ’21). Association for Computing Machinery, New York, NY, USA, Article 163, 13 pages. doi:10.1145/3411764.3445231 Burgoon and Le Poire (1993) Judee K Burgoon and Beth A Le Poire. 1993. Effects of communication expectancies, actual communication, and expectancy disconfirmation on evaluations of communicators and their communication behavior. Human communication research 20, 1 (1993), 67–96. Cassell (2001) Justine Cassell. 2001. Embodied Conversational Agents: Representation and Intelligence in User Interfaces. AI Magazine 22, 4 (2001), 67–83. Cho et al. (2017) Hee-Kyung Cho, Matthias Trier, and Eunhee Kim. 2017. The Use of Instant Messaging in Working Relationship Development: a Case Study. Journal of Computer-Mediated Communication 10, 4 (07 2017), JCMC1044. doi:10.1111/j.1083-6101.2005.tb00280.x Clark et al. (2019) Leigh Clark, Nadia Pantidi, Orla Cooney, Philip Doyle, Diego Garaialde, Justin Edwards, Brendan Spillane, Emer Gilmartin, Christine Murad, Cosmin Munteanu, Vincent Wade, and Benjamin R. Cowan. 2019. What Makes a Good Conversation? Challenges in Designing Truly Conversational Agents. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (Glasgow, Scotland Uk) (CHI ’19). Association for Computing Machinery, New York, NY, USA, 1–12. doi:10.1145/3290605.3300705 Crawford and Hamidi (2025) Kirk Andrew Crawford and Foad Hamidi. 2025. ”Like a Love Language”: Understanding Communication in Disabled LGBTQIA+ Romantic Relationships. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 403, 17 pages. doi:10.1145/3706598.3713195 Cui et al. (2013) Guoqiang Cui, Barbara Lockee, and Cuiqing Meng. 2013. Building modern online social presence: A review of social presence theory and its instructional design implications for future trends. Education and information technologies 18, 4 (2013), 661–685. Falbo and Peplau (1980) Toni Falbo and Letitia A Peplau. 1980. Power strategies in intimate relationships. Journal of personality and social psychology 38, 4 (1980), 618. Fattorini (1968) Hector O Fattorini. 1968. Boundary control systems. SIAM Journal on Control 6, 3 (1968), 349–385. Fay et al. (2018) Nicolas Fay, Bradley Walker, Nik Swoboda, and Simon Garrod. 2018. How to create shared symbols. Cognitive Science 42 (2018), 241–269. Fourie et al. (2022) Christopher Fourie, Nadia Figueroa, Julie Shah, Marta Bieńkiewicz, Benoît Bardy, Etienne Burdet, Phani Teja Singamaneni, Rachid Alami, Arianna Curioni, Günther Knoblich, Wafa Johal, Dagmar Sternad, and Malte Jung. 2022. Joint Action, Adaptation, and Entrainment in Human-Robot Interaction. In Proceedings of the 2022 ACM/IEEE International Conference on Human-Robot Interaction (Sapporo, Hokkaido, Japan) (HRI ’22). IEEE Press, 1250–1253. Fowler (1987) Loretta Fowler. 1987. Shared symbols, contested meanings: Gros Ventre culture and history, 1778-1984. Cornell University Press. Furman and Wehner (1997) Wyndol Furman and Elizabeth A Wehner. 1997. Adolescent romantic relationships: A developmental perspective. New directions for child and adolescent development 1997, 78 (1997), 21–36. Griggio et al. (2019) Carla F. Griggio, Joanna McGrenere, and Wendy E. Mackay. 2019. Customizations and Expression Breakdowns in Ecosystems of Communication Apps. Proc. ACM Hum.-Comput. Interact. 3, CSCW, Article 26 (Nov. 2019), 26 pages. doi:10.1145/3359128 Griggio et al. (2021) Carla F. Griggio, Arissa J. Sato, Wendy E. Mackay, and Koji Yatani. 2021. Mediating Intimacy with DearBoard: a Co-Customizable Keyboard for Everyday Messaging. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (Yokohama, Japan) (CHI ’21). Association for Computing Machinery, New York, NY, USA, Article 342, 16 pages. doi:10.1145/3411764.3445757 Groffman et al. (2010) Peter M Groffman, Cathlyn Stylinski, Matthew C Nisbet, Carlos M Duarte, Rebecca Jordan, Amy Burgin, M Andrea Previtali, and James Coloso. 2010. Restarting the conversation: challenges at the interface between ecology and society. Frontiers in Ecology and the Environment 8, 6 (2010), 284–291. Gross and John (1998) James J Gross and Oliver P John. 1998. Mapping the domain of expressivity: multimethod evidence for a hierarchical model. Journal of personality and social psychology 74, 1 (1998), 170. Hanks (2013) William F Hanks. 2013. Counterparts: Co-presence and ritual intersubjectivity. Language & Communication 33, 3 (2013), 263–277. Hart (2006) Sandra G Hart. 2006. NASA-task load index (NASA-TLX); 20 years later. In Proceedings of the human factors and ergonomics society annual meeting, Vol. 50. Sage publications Sage CA: Los Angeles, CA, 904–908. Hassenzahl et al. (2012) Marc Hassenzahl, Stephanie Heidecker, Kai Eckoldt, Sarah Diefenbach, and Uwe Hillmann. 2012. All You Need is Love: Current Strategies of Mediating Intimate Relationships through Technology. ACM Trans. Comput.-Hum. Interact. 19, 4, Article 30 (Dec. 2012), 19 pages. doi:10.1145/2395131.2395137 Hertenstein et al. (2009) Matthew J Hertenstein, Rachel Holmes, Margaret McCullough, and Dacher Keltner. 2009. The communication of emotion via touch. Emotion 9, 4 (2009), 566. Homaeian et al. (2021) Leila Homaeian, James R. Wallace, and Stacey D. Scott. 2021. Joint Action Storyboards: A Framework for Visualizing Communication Grounding Costs. Proc. ACM Hum.-Comput. Interact. 5, CSCW1, Article 28 (April 2021), 27 pages. doi:10.1145/3449102 Hübler (2011) Axel Hübler. 2011. Communication and expressivity. In Functionalism in linguistics. John Benjamins Publishing Company, 357–380. Ichino et al. (2022) Junko Ichino, Masahiro Ide, Hitomi Yokoyama, Hirotoshi Asano, Hideo Miyachi, and Daisuke Okabe. 2022. ”I’ve talked without intending to”: Self-disclosure and Reciprocity via Embodied Avatar. Proc. ACM Hum.-Comput. Interact. 6, CSCW2, Article 482 (Nov. 2022), 23 pages. doi:10.1145/3555583 Ikeya et al. (2024) Yuta Ikeya, Miguel Bruns, and Bahareh Barati. 2024. Designing for Interdependence of Bees, Garden, Designer, and the Changing Season. In Proceedings of the 2024 ACM Designing Interactive Systems Conference (Copenhagen, Denmark) (DIS ’24). Association for Computing Machinery, New York, NY, USA, 1385–1399. doi:10.1145/3643834.3660743 Jones et al. (2011) Tim Jones, Peter A Dacin, and Shirley F Taylor. 2011. Relational damage and relationship repair: A new look at transgressions in service relationships. Journal of Service Research 14, 3 (2011), 318–339. Keltner et al. (2019) Dacher Keltner, Disa Sauter, Jessica Tracy, and Alan Cowen. 2019. Emotional expression: Advances in basic emotion theory. Journal of nonverbal behavior 43, 2 (2019), 133–160. Kianzad and Seifi (2024) Soheil Kianzad and Hasti Seifi. 2024. MagicDraw: Haptic-Assisted One-Line Drawing with Shared Control. In Adjunct Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology (Pittsburgh, PA, USA) (UIST Adjunct ’24). Association for Computing Machinery, New York, NY, USA, Article 27, 3 pages. doi:10.1145/3672539.3686753 Kim et al. (2023) Jieun Kim, Gonzalo Gonzalez-Pumariega, Soyee Park, and Susan R. Fussell. 2023. Urgency Builds Trust: A Voice Agent’s Emotional Expression in an Emergency. In Companion Publication of the 2023 Conference on Computer Supported Cooperative Work and Social Computing (Minneapolis, MN, USA) (CSCW ’23 Companion). Association for Computing Machinery, New York, NY, USA, 343–347. doi:10.1145/3584931.3606979 Knierim et al. (2025) Pascal Knierim, Jacob Höck, and Thomas Kosch. 2025. Wrist-Powered Touch: Evaluating Smartwatch-Based Touch Gesture Recognition for Interaction in Extended Reality. In Proceedings of the Mensch Und Computer 2025 (MuC ’25). Association for Computing Machinery, New York, NY, USA, 345–353. doi:10.1145/3743049.3743090 Kong et al. (2025) Ha-Kyung Kong, Rachel Lowy, Youjin Choi, and Jennifer G Kim. 2025. Working Together Toward Interdependence: Chatbot-Based Support for Balanced Social Interactions Between Neurodivergent and Neurotypical Individuals. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 779, 17 pages. doi:10.1145/3706598.3713344 Kreijns et al. (2022) Karel Kreijns, Ke Xu, and Joshua Weidlich. 2022. Social Presence: Conceptualization and Measurement. Educational Psychology Review 34 (2022), 139–170. doi:10.1007/s10648-021-09623-8 Larsen and Johnson (2017) Soren C Larsen and Jay T Johnson. 2017. Being together in place: Indigenous coexistence in a more than human world. U of Minnesota Press. Li et al. (2018) Xiaomin Li, Hongjian Cao, Nan Zhou, Xiaoyan Ju, Jing Lan, Qinyi Zhu, and Xiaoyi Fang. 2018. Daily communication, conflict resolution, and marital quality in Chinese marriage: A three-wave, cross-lagged analysis. Journal of Family Psychology 32, 6 (2018), 733. Li et al. (2022) Yanhong Li, Zhenhan Gao, Sabrina Egger, Sven Mayer, and Heinrich Hussmann. 2022. Tangible Interfaces Support Young Children’s Goal Interdependence. In Proceedings of Mensch Und Computer 2022 (Darmstadt, Germany) (MuC ’22). Association for Computing Machinery, New York, NY, USA, 147–157. doi:10.1145/3543758.3543782 Liao et al. (2024) Zimo Liao, Zhicheng Luo, Qianyi Huang, Linfeng Zhang, Fan Wu, Qian Zhang, and Guihai Chen. 2024. Gesture Recognition Using Visible Light on Mobile Devices. IEEE/ACM Trans. Netw. 32, 4 (March 2024), 2920–2935. doi:10.1109/TNET.2024.3369996 Ling (2010) Rich Ling. 2010. New Tech, New Ties: How Mobile Communication Is Reshaping Social Cohesion. MIT Press, Cambridge, MA. Liu (2024) Yuexuan Liu. 2024. Development of Wearable Medical Devices in Parkinson’s Disease. In Proceedings of the 2024 International Conference on Smart Healthcare and Wearable Intelligent Devices (SHWID ’24). Association for Computing Machinery, New York, NY, USA, 156–160. doi:10.1145/3703847.3703874 Miyata and Osawa (2024) Kansei Miyata and Hirotaka Osawa. 2024. Enhancing Cooperative Behavior through Subtle Communication Cues: The Impact of Hint Order and Hesitation in Hanabi. In Proceedings of the 12th International Conference on Human-Agent Interaction (Swansea, United Kingdom) (HAI ’24). Association for Computing Machinery, New York, NY, USA, 444–446. doi:10.1145/3687272.3690912 Morgan (1943) Clifford Thomas Morgan. 1943. Physiological psychology. (1943). Nardi et al. (2000) Bonnie A. Nardi, Steve Whittaker, and Erin Bradner. 2000. Interaction and outeraction: instant messaging in action. In Proceedings of the 2000 ACM Conference on Computer Supported Cooperative Work (Philadelphia, Pennsylvania, USA) (CSCW ’00). Association for Computing Machinery, New York, NY, USA, 79–88. doi:10.1145/358916.358975 Oh et al. (2018) Catherine S Oh, Jeremy N Bailenson, and Gregory F Welch. 2018. A systematic review of social presence: Definition, antecedents, and implications. Frontiers in Robotics and AI 5 (2018), 114. Overall et al. (2009) Nickola C Overall, Garth JO Fletcher, Jeffry A Simpson, and Chris G Sibley. 2009. Regulating partners in intimate relationships: the costs and benefits of different communication strategies. Journal of personality and social psychology 96, 3 (2009), 620. Panda et al. (2025) Abhinandan Panda, Srinivas Pinisetty, and Partha Roop. 2025. Securing Pacemakers Using Runtime Monitors over Physiological Signals. ACM Transactions on Embedded Computing Systems 24, 3, Article 50 (May 2025), 41 pages. doi:10.1145/3638286 Pettys-Baker et al. (2024) Robert Pettys-Baker, Megan E. Clarke, and Brad Holschuh. 2024. Functional Now, Wearable Later: Examining the Design Practices of Wearable Technologists. In Proceedings of the 2024 ACM International Symposium on Wearable Computers (Melbourne VIC, Australia) (ISWC ’24). Association for Computing Machinery, New York, NY, USA, 71–81. doi:10.1145/3675095.3676615 Podkosova et al. (2023) Iana Podkosova, Francesco De Pace, and Hugo Brument. 2023. Joint Action in Collaborative Mixed Reality: Effects of Immersion Type and Physical Location. In Proceedings of the 2023 ACM Symposium on Spatial User Interaction (Sydney, NSW, Australia) (SUI ’23). Association for Computing Machinery, New York, NY, USA, Article 16, 12 pages. doi:10.1145/3607822.3614541 Reid et al. (1996) Fraser JM Reid, Vlastimil Malinek, Clifford JT Stott, and Jonathan St BT Evans. 1996. The messaging threshold in computer-mediated communication. Ergonomics 39, 8 (1996), 1017–1037. Riordan and Kreuz (2010) Monica A. Riordan and Roger J. Kreuz. 2010. Cues in computer-mediated communication: A corpus analysis. Computers in Human Behavior 26, 6 (2010), 1806–1817. doi:10.1016/j.chb.2010.07.008 Online Interactivity: Role of Technology in Behavior Change. Roediger and Pyc (2012) Henry L. Roediger and Mary A. Pyc. 2012. Inexpensive techniques to improve education: Applying cognitive psychology to enhance educational practice. Journal of Applied Research in Memory and Cognition 1, 4 (2012), 242–248. doi:10.1016/j.jarmac.2012.09.002 Rosenthal (1967) Robert Rosenthal. 1967. Unintended communication of interpersonal expectations. American Behavioral Scientist 10, 8 (1967), 24–26. Roto et al. (2016) Virpi Roto, Heli Vätäjä, Effie Law, and Rachel Powers. 2016. Experience Design for Multiple Customer Touchpoints. In Proceedings of the 9th Nordic Conference on Human-Computer Interaction (Gothenburg, Sweden) (NordiCHI ’16). Association for Computing Machinery, New York, NY, USA, Article 146, 3 pages. doi:10.1145/2971485.2987685 Rusbult and Van Lange (2008) Caryl E Rusbult and Paul AM Van Lange. 2008. Why we need interdependence theory. Social and personality psychology compass 2, 5 (2008), 2049–2070. Sawmong (2022) Sudaporn Sawmong. 2022. The role of personality traits and engagement factors towards the use of paid stickers in personal communication messages. F1000Research 11 (2022), 915. Schwarz (2011) Ori Schwarz. 2011. Who moved my conversation? Instant messaging, intertextuality and new regimes of intimacy and truth. Media, Culture & Society 33, 1 (2011), 71–87. Shi et al. (2024) Yunzi Shi, John Bell, and Adithya Pediredla. 2024. Embodied Visions: Interactive Installations That Reimagine Bodily Presence in Digital Imaging Apparatuses as Shadows. In SIGGRAPH Asia 2024 Art Papers (SA ’24). Association for Computing Machinery, New York, NY, USA, Article 8, 8 pages. doi:10.1145/3680530.3695436 Simpson (1990) Jeffry A Simpson. 1990. Influence of attachment styles on romantic relationships. Journal of personality and social psychology 59, 5 (1990), 971. Skeggs et al. (2025) Amira Skeggs, Ashish Mehta, Valerie Yap, Seray B Ibrahim, Charla Rhodes, James J. Gross, Sean A. Munson, Predrag Klasnja, Amy Orben, and Petr Slovak. 2025. Micro-narratives: A Scalable Method for Eliciting Stories of People’s Lived Experience. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 330, 20 pages. doi:10.1145/3706598.3713999 Sproull and Kiesler (1986) Lee Sproull and Sara Kiesler. 1986. Reducing social context cues: Electronic mail in organizational communication. Management science 32, 11 (1986), 1492–1512. Ståhl et al. (2005) Anna Ståhl, Kristina Hök, and Petra Sundström. 2005. A foundation for emotional expressivity. In DUX, Designing for User Experience. AIGA. Starke et al. (2024) Sebastian Starke, Paul Starke, Nicky He, Taku Komura, and Yuting Ye. 2024. Categorical Codebook Matching for Embodied Character Controllers. ACM Trans. Graph. 43, 4, Article 142 (July 2024), 14 pages. doi:10.1145/3658209 Stone et al. (2005) Debbie Stone, Caroline Jarrett, Mark Woodroffe, and Shailey Minocha. 2005. User interface design and evaluation. Elsevier. Swallow and Jiang (2013) Khena M Swallow and Yuhong V Jiang. 2013. Attentional load and attentional boost: A review of data and theory. Frontiers in Psychology 4 (2013), 274. Tang et al. (2024) Xinyao Tang, Xinyi Zhang, Xupeng Wang, Xinran Cheng, Hongyan Liu, and Haonan Jin. 2024. Design and Research of a Wearable Portable Electromyography Monitoring System. In Proceedings of the 2024 International Conference on Smart Healthcare and Wearable Intelligent Devices (SHWID ’24). Association for Computing Machinery, New York, NY, USA, 192–198. doi:10.1145/3703847.3703880 Tang and Hew (2019) Ying Tang and Khe Foon Hew. 2019. Emoticon, emoji, and sticker use in computer-mediated communication: A review of theories and research findings. International journal of communication 13 (2019), 27–27. Vasseleu (1999) Cathryn Vasseleu. 1999. Touch, digital communication and the ticklish. Angelaki: journal of the theoretical humanities 4, 2 (1999), 153–162. Venkatraj et al. (2025) Karthikeya Puttur Venkatraj, Christina Schneegass, Gijs Huisman, and Abdallah El Ali. 2025. Demonstrating Dark Haptics Scenarios for Future XR Advertising. In Adjunct Proceedings of the 4th Annual Symposium on Human-Computer Interaction for Work (CHIWORK ’25 Adjunct). Association for Computing Machinery, New York, NY, USA, Article 20, 3 pages. doi:10.1145/3707640.3731926 Verheijen (2013) Lieke Verheijen. 2013. The effects of text messaging and instant messaging on literacy. English studies 94, 5 (2013), 582–602. Wageman (2014) Ruth Wageman. 2014. The meaning of interdependence. In Groups at work. Psychology Press, 197–217. Wagener et al. (2025) Nadine Wagener, Daniel Christian Albensoeder, Leon Reicherts, Paweł W. Woźniak, Yvonne Rogers, and Jasmin Niess. 2025. TogetherReflect: Supporting Emotional Expression in Couples Through a Collaborative Virtual Reality Experience. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 1130, 16 pages. doi:10.1145/3706598.3713642 Walther (1992) Joseph B. Walther. 1992. Interpersonal effects in computer-mediated interaction: A relational perspective. Communication Research 19, 1 (1992), 52–90. Walther (1996) Joseph B. Walther. 1996. Computer-mediated communication: Impersonal, interpersonal, and hyperpersonal interaction. Communication Research 23, 1 (1996), 3–43. Wang et al. (2023) Xian Wang, Xiaoyu Mo, Lik-Hang Lee, Xiaoying Wei, Xiaofu Jin, Mingming Fan, and Pan Hui. 2023. Designing Loving-Kindness Meditation in Virtual Reality for Long-Distance Romantic Relationships. In Proceedings of the 31st ACM International Conference on Multimedia (Ottawa ON, Canada) (M ’23). Association for Computing Machinery, New York, NY, USA, 7608–7617. doi:10.1145/3581783.3612438 Wijesinghe et al. (2025) Nipuni H. Wijesinghe, Yue Peng, Sharni Konrad, Maleen Jayasuriya, Janie Busby Grant, and Damith Herath. 2025. Reframing Social Presence for Human Robot Interaction. In Proceedings of the 2025 ACM/IEEE International Conference on Human-Robot Interaction (Melbourne, Australia) (HRI ’25). IEEE Press, 1716–1721. Woerner et al. (2007) Stephanie L. Woerner, JoAnne Yates, and Wanda J. Orlikowski. 2007. Conversational Coherence in Instant Messaging and Getting Work Done. In 2007 40th Annual Hawaii International Conference on System Sciences (HICSS’07). 77–77. doi:10.1109/HICSS.2007.152 Zhu et al. (2025) Jingwen Zhu, Fiona Bell, Katherine W Song, Katia Vega, Aditya Shekhar Nittala, Mirela Alistar, Leah Buechley, and Cindy Hsin-Liu Kao. 2025. Wearable Bio-HCI: Challenges & Opportunities. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA ’25). Association for Computing Machinery, New York, NY, USA, Article 866, 5 pages. doi:10.1145/3706599.3716295 Appendix A Questionnaire To assess participants’ subjective experiences with the system, we developed a questionnaire based on four conceptual subscales reflecting key interaction qualities. Each subscale consists of multiple facets capturing distinct aspects of the experience, with both positively and negatively worded items included. The full set of items is summarized in Table 5. Table 5. BASIC scale items by subscale and facet. (R) indicates reverse coded items. BASIC scale items by subscale and facet. (R) indicates reverse-coded items. Subscale Facet Item text Co-presence and Affective Intimacy Presence When using PuppetChat, I often feel as if my partner is right in front of me. Disclosure In conversations, I am more willing to share my genuine feelings or updates. Rituals We developed some expressions or rituals unique to our relationship. Warmth Compared to our usual chatting style, our relationship feels warmer. Reciprocal Interaction Pairs We often exchange paired or complementary actions. Understanding I can easily understand the intention behind my partner’s actions and respond accordingly. Fit The reciprocal actions recommended by the system are usually appropriate and easy to use. Rhythm The combinations and matches of interactive actions make our conversations more rhythmic and enjoyable. Personalized Micronarratives Connection The system generated micronarratives often naturally connect to my past experiences. Story These micronarratives help us link fragmented conversations into more continuous little stories. Reuse I would like to reuse previous micronarratives as a form of continuity. Empty (R) Some micronarratives feel empty, like templates. Agency and Persistence Editing I can easily edit or rewrite the actions or micronarratives recommended by the system. Control This editing ability makes me feel more comfortable and in control. Memory For meaningful or memorable interactions, I prefer that they can be saved and revisited later. Ephemeral (R) I prefer that all expressions quickly disappear without leaving any record. Appendix B Evaluation We conducted a user study with 11 dyads to evaluate the system in realistic interpersonal communication contexts. Participants varied in age, relationship type, occupation, and primary messaging platforms, providing a diverse sample of everyday messaging practices. Detailed participant demographics and background information are shown in Table 6. Table 6. Participants (11 dyads): demographics, relationship, occupation, and messaging apps Participants (11 dyads): demographics, relationship, occupation, and messaging apps ID Gender Age Relationship Occupation Messaging Apps P1-A Female 24 Best friends Student WeChat; WhatsApp; Instagram P1-B Female 23 Best friends Student P2-A Female 25 Couple HR Business Partner WeChat; WhatsApp P2-B Male 26 Couple Software Developer P3-A Female 25 Best friends Student LINE; Instagram P3-B Female 22 Best friends Student P4-A Female 24 Couple Data Analyst iMessage; WhatsApp P4-B Male 26 Couple Bank Staff P5-A Male 23 Best friends Student Discord; Telegram P5-B Male 23 Best friends Student P6-A Female 27 Couple Nurse WhatsApp; Signal P6-B Male 27 Couple Civil Engineer P7-A Female 21 Best friends student Snapchat; Instagram P7-B Female 21 Best friends student P8-A Male 28 Couple Software Engineer WeChat; Telegram P8-B Female 27 Couple Software Engineer P9-A Male 26 Best friends Software Engineer Instagram; Signal; Twitter/X P9-B Male 25 Best friends Data Scientist P10-A Female 23 Couple Research Assistant WhatsApp; Facebook Messenger P10-B Male 24 Couple Student P11-A Male 24 Best friends High School Teacher LINE; Telegram P11-B Female 24 Best friends Marketing Specialist

Instant research discovery