Source-linked AI summary
Designing Proactive Thought Partners for Writing
Chao Zhang, Abe Davis, Chih-Wei Chen, Chin-Chia Hsu
TL;DR
Writing support must adapt to changing cognitive needs, while existing proactive tools have focused largely on local textual assistance. This paper studies customizable proactive thought partners through a technology probe deployed with 16 participants for one week, finding that users planned support prospectively and valued graduated, lightweight interventions aligned with their writing moments.
Problem
Existing proactive writing assistance largely provides local textual continuations, leaving the design of functional systems for customizable, context-sensitive higher-level cognitive support unclear.
Method
The authors built a technology probe that lets writers configure partners’ cognitive roles and proactivity, then deployed it in a one-week diary study with 16 participants.
Results
Participants used configuration as prospective planning, engaged suggestions for idea generation and self-monitoring, and found lightweight visual and non-directive rhetorical representations less intrusive.
Takeaways & Limitations
Proactive writing assistants should support customization, contextual timing, graduated engagement, and non-intrusive visual and rhetorical representations.
Takeaways & Limitations
As a probe study with English-proficient daily writers experienced with AI tools, the findings do not establish causal effects and may not generalize broadly.
Abstract
from arXiv · showhide
Writing involves diverse cognitive activities, from ideation to revision, and writers' needs vary across individuals and moments. Proactive AI promises to provide the right support at the right time, yet existing proactive tools largely focus on generic textual assistance, such as autocomplete. This paper studies the design space of proactive thought partners: AI agents that proactively offer customizable, higher-level cognitive support during writing. We instantiated this concept in a technology probe and deployed it with 16 participants for one week. The probe allows users to create partners by configuring their roles and proactivity. As users write, relevant partners take the initiative at appropriate moments to offer suggestions. Our findings show that participants configured proactive support through prospective planning, used suggestions for both idea generation and self-monitoring, and valued lightweight visual representations alongside non-directive rhetorical framing for non-intrusive interventions. We derive implications for designing proactive writing assistants around customization, timing, engagement, and representation.
1. Introduction
Writing demands changing forms of cognitive support across moments and writers, while existing proactive assistance mainly offers local textual continuations. This paper introduces customizable proactive thought partners for higher-level cognitive support during writing and studies their use through a one-week probe deployment.
- Motivation: Writing shifts among developing ideas, selecting prose, and evaluating text against broader narrative goals, creating changing support needs.Writers may need examples, alternative perspectives, or help developing a next section at different moments.
- Motivation: Existing proactive writing tools primarily use surrounding text and interaction events to provide local assistance such as autocomplete and next-phrase suggestions.These approaches address relatively structured text production rather than higher-level personal cognitive needs.
- Concept: Proactive thought partners provide customizable, higher-level cognitive support without requiring explicit user prompting.The concept combines system initiative with user-defined cognitive support during writing.
- Probe: The technology probe lets writers configure partners’ roles and timing conditions, then dismiss, use, or apply suggestions directly while writing.Users define what support a partner provides and when it may intervene.
- Findings: In a one-week diary study with 16 participants, users treated configuration as prospective planning and used suggestions for idea generation and self-monitoring.Participants also valued lightweight visual representations and non-directive rhetorical framing for less intrusive interventions.
2. Related Work
Related work establishes a progression from textual writing assistance toward proactive and customizable cognitive support. The paper addresses the unresolved design of functional systems that interpret writing context and let users shape when cognitive assistance arrives.
- Intelligent writing tools: Intelligent writing tools support planning, translation, and review through production, correction, and refinement assistance.Traditional systems include autocomplete and predictive text for word- or phrase-level continuations.
- Intelligent writing tools: LLM-based writing systems broaden assistance to brainstorming, outlining, drafting, rewriting, and feedback.This expands support beyond correcting or completing text.
- Proactive AI assistants: Proactive AI infers user needs and offers assistance without requiring an explicit prompt, extending mixed-initiative interaction.Prior writing systems commonly tie intervention timing to observable events such as typing or pauses.
- Proactive AI assistants: A Wizard-of-Oz study found simulated proactive support could enhance creativity, but functional-system design, contextual timing, and user customization remained unclear.This paper instantiates proactive cognitive support in a functional system with customizable thought partners.
- Customizable AI agents: Existing customizable writing agents are primarily invoked on demand, whereas this work adds customization of the AI agent’s proactivity.Writers shape both the cognitive support they receive and when it enters the writing process.
3. Design Goals
The technology probe is designed to investigate an open-ended design space through writers’ own tasks rather than evaluate an optimal finished system. Its goals prioritize configurable support, flexible agency-preserving engagement, and lightweight contextual integration.
- Probe rationale: The technology probe provides a functional, deployable writing environment for investigating users’ practices, needs, values, and experiences with proactive thought partners.Its purpose is exploration rather than optimization of a finished writing system.
- Design goals: Users configure both partner roles and proactivity so support can reflect expected writing needs and preferred intervention conditions.Role specifies the kind of support; proactivity specifies when it may intervene.
- Design goals: The probe preserves agency by supporting different levels of AI involvement, from inspiration to direct text generation.This accommodates moments when writers may or may not want assistance.
- Design goals: Proactive support is embedded in a lightweight writing environment that supports ongoing composition and revision without making the interface the study’s primary focus.The environment also monitors writing traces, detects candidate intervention moments, and surfaces support in context.
- Design goals: The writing environment combines a lightweight Markdown editor with a side panel that surfaces contextually relevant partner suggestions alongside the draft.The interface supports proactive assistance during writers’ ongoing tasks.
4. Proactive Thought Partners: A Technology Probe
The probe lets writers define partner roles, event triggers, and contextual heuristics, then uses writing context to activate relevant partners. Suggestions support graduated engagement, from unobtrusive ignoring through inspiration and optional direct writing actions.
- Probe interfaces: The probe includes onboarding, partner configuration, and a lightweight Markdown editor for writing with enabled partners.These interfaces support session goals, partner creation, and situated writing practice.
- Partner customization: Each partner combines a role, event triggers, and a contextual heuristic to specify what support it provides and when it is relevant.Triggers identify broad candidate moments, while heuristics refine relevance to the current writing context.
- Partner activation: The system considers pauses, sentence ends, and text selection as candidate intervention events, with customizable idle thresholds for pauses and selections.Sentence-end activation waits for a short idle period to avoid interrupting rapid typing.
- Partner activation: After a trigger, the decision engine evaluates session goals, current text, recent writing behaviors, and enabled-partner heuristics before selecting at most two partners.Writing behaviors include cursor position, trigger event, and keystroke logs from the preceding 15 seconds.
- Suggestion generation: Each suggestion acknowledges inferred writer intent and offers a role-tailored question-style prompt.The acknowledgment locates the suggestion in the writer’s current activity, while the question framing supports thought-provoking assistance.
- Engagement: Engagement progresses from ignoring a fading tag, to inspecting or discussing inspiration, to selectively executing text changes while retaining control.Writers can accept or revert direct insertions or revisions, preserving agency across increasing AI involvement.
5. User Study
The study used an exploratory technology-probe deployment to examine how writers configure and engage with proactive thought partners in open-ended writing practice.
- Study Design: 16 participants used the probe in a one-week diary study followed by semi-structured interviews.The study combined interaction logs, diary entries, and interview data to examine partner creation, suggestion engagement, and user experience.
- Study Design: The technology probe was designed to explore adaptation and appropriation of a nascent technology rather than evaluate the efficacy of a finished product.This framing guided the deployment toward uncovering how users made sense of proactive thought partners in practice.
- Participants: Participants had prior experience using AI tools for writing, with 10 reporting daily use and 6 reporting weekly use before the study.The sample consisted of everyday writers who regularly used AI-assisted writing for activities including brainstorming, drafting, and editing.
- Study Design: Participants completed open-ended writing sessions across personal, academic, professional, creative, technical, and journalistic contexts.They were asked to complete at least four substantive sessions, defined by active writing duration or word production and completion of a self-contained piece.
- Data Collection: The probe collected partner configurations, keystroke and interaction logs, binary intervention feedback, diary ratings, and open-ended reflections.Daily diaries measured satisfaction, perceived timeliness, and perceived helpfulness alongside qualitative accounts of configured partners and received support.
6. Findings
Participants configured proactive thought partners around anticipated writing needs and used their suggestions selectively for generating ideas, monitoring progress, and executing text. They generally valued support that aligned with current intentions while preserving flow, control, and low-disruption interaction.
- Partner Creation: Participants configured partners prospectively by anticipating writing goals and difficulties, then specifying relevant roles and timing conditions.Event triggers served as broad candidate moments, while contextual heuristics captured draft state, writing activity, and anticipated support needs.
- Partner Creation: 43 of 54 partners (79.63%) provided higher-level cognitive support, while 11 (20.37%) provided local textual assistance.Roles included information seeking, argument development, critical reflection, and ideation support.
- Partner Creation: Participants configured event triggers broadly: pauses and sentence endings were each enabled for 40 of 54 partners (74.07%), while text selections were enabled for 44 (81.48%).Contextual heuristics narrowed when support was relevant, including draft-state, writing-activity, and anticipated-need cues.
- Suggestion Engagement: Suggestions supported both idea generation and self-monitoring, introducing unexpected connections while helping writers detect drift, gaps, or neglected perspectives.Writers used suggestions as prompts for extending their thinking and checking alignment with their goals and arguments.
- Suggestion Engagement: 58.27% of 1,100 interventions were ignored, especially after pauses, which participants used to preserve writing flow when they already knew what to write.Pause-triggered interventions were ignored 64.39% of the time, reflecting ambiguity between cognitive need and focused concentration.
- Suggestion Engagement: 22.18% of interventions led to execution, and 84.43% of generated text was accepted; execution was most common after text selection and when writers were ready to externalize a developed intention.Writers also executed out of curiosity or to translate a clear idea into fluent text, while reflection-oriented suggestions were less likely to be executed.
7. Discussion
The discussion frames proactive thought partners as customizable writing assistants whose usefulness depends on prospective planning, contextual timing, graduated engagement, and lightweight representation.
- Design concept: Proactive thought partners combine system initiative with customizable, higher-level cognitive support for writing.The design concept is explored through customization, timing, engagement, and representation.
- Customization as prospective planning: Customization lets writers translate anticipated goals and difficulties into specialized partners for different forms of cognitive work.Configuration externalizes a situated model of forthcoming writing work and can support information seeking, argument development, critical reflection, and ideation.
- Customization as prospective planning: Partner configuration should be iterative because writers cannot always predict which roles and timing conditions will help once writing unfolds.Systems could support previews, simulations, and revisions based on use.
- Timing as contextual alignment: Observable events such as pauses and sentence endings should trigger contextual assessment rather than determine intervention on their own.The same event can reflect different writing states, so systems should interpret it alongside draft content, activity, and anticipated needs.
- Engagement as graduated commitment: Engagement ranges from effortless ignoring to inspiration or reflection and selective direct execution.Ignoring preserves flow, while execution is most appropriate when writers have a sufficiently settled intention.
- Representation: Peripheral placement, fading presence, acknowledgments, and questions make proactive suggestions noticeable while preserving writer judgment.These representations frame interventions as optional bids for attention that can recede without competing with the writing task.
8. Ethical Considerations
Proactive writing support creates privacy risks because timely assistance requires continuous observation of drafts and interaction traces.
- Privacy risks: Continuous monitoring can expose sensitive document content and distinctive writing habits that may reveal aspects of a writer’s identity.The probe observed the evolving draft, cursor position, behavioral events, and recent keystroke logs.
- Privacy safeguards: Privacy-by-design measures should include data minimization, local or ephemeral processing, retention limits, monitoring disclosure, and controls to pause or delete data.Systems should not reuse writing traces for training, profiling, or identity inference without separate explicit consent.
9. Limitations
The probe offers exploratory evidence rather than causal evaluation, and its findings may not generalize across populations, contexts, models, or longer-term use.
- Study scope: As a probe study, the research does not establish causal effects on writing quality, productivity, learning, or long-term agency.Longer deployments and comparative quantitative studies are suggested for future work.
- Generalizability: The participant sample comprised English-proficient daily writers with prior AI-writing-tool experience, limiting generalizability to other writers and contexts.The privacy instruction also likely excluded high-stakes workplace writing.
- Context inference: Current LLMs may not reliably infer writers’ momentary intentions or anticipated support needs from contextual cues.Future work should evaluate recognition of draft-state, writing-activity, and anticipated-need cues across different writing states.
- Implementation dependence: Suggestion quality and timing depended on one LLM-based implementation, so other models, prompts, representations, or orchestration strategies may produce different experiences.The findings reflect both the broader interaction paradigm and the implemented system’s capabilities and limitations.
10. Conclusion
The paper introduces proactive thought partners for customizable, higher-level cognitive support during writing. A one-week probe shows that effective proactivity combines prospective planning, graduated engagement, and contextual alignment with writers’ momentary intentions.
- Contribution: Proactive thought partners are AI writing assistants that proactively provide customizable, higher-level cognitive support during writing.The concept combines system initiative with support shaped by writers’ needs.
- Findings: A one-week probe found that writers used customization as prospective planning and engaged with suggestions through graduated levels of commitment.Good timing was experienced as contextual alignment between offered support and momentary intentions.
- Implication: Effective proactivity requires writers to shape support, ignore or deepen engagement with suggestions, and receive interventions aligned with their current intentions.The conclusion emphasizes that deciding when to intervene is not sufficient by itself.