Source-linked AI summary
What Makes an Initial Reaction Ready for Discussion?: Multi-Persona AI Support for Stance Reflection and Writing
Sky Shih-Kai Hong, Mu-Tien Kuo, Wei-Ji Chen, Dennis Wang
TL;DR
People may need to clarify claims, anticipate audience challenges, and choose what reasoning to reveal before sharing a social or community stance. StanceLab compares parallel Interviewer, Mentor, and Opponent responses with a standalone LLM in a formative pilot, finding that useful diagnoses require coordination support for turning reflection into selective final messages.
Problem
Preparing a stance requires clarification, organization, challenge, and audience-aware disclosure before an initial reaction becomes a shareable message.
Method
StanceLab compares a three-persona mode with a standalone LLM in a formative within-subject pilot using six participants across two tasks each.
Results
All 12 sessions produced short final notepad messages, while useful persona moments surfaced actionable blind spots, questions, objections, or reframes.
Takeaways & Limitations
Persona diversity is most promising as a diagnostic layer connected to writing support that preserves personal voice and audience-appropriate disclosure choices.
Takeaways & Limitations
The small formative pilot drew from a student developer community, varied topics and audiences, and recorded post-session confidence in only 9 of 12 sessions.
Abstract
from arXiv · showhide
An initial reaction to a social or community issue can feel meaningful before it is ready to become a message: people still need to clarify the claim, anticipate audience risks, and decide how much reasoning should become visible to others. We present StanceLab, a prototype for preparing a stance before entering a discussion. The prototype compares a three-persona mode, where an Interviewer, Mentor, and Opponent respond in parallel to help users diagnose and revise a stance, with a standalone LLM mode. In a formative within-subject pilot with six participants and 12 task sessions, every session produced a short final message in the notepad. The pilot revealed two design requirements: persona roles should diagnose useful blind spots or objections, and parallel responses need coordination support. We propose a future diagnosis-and-writing workflow that turns persona-based reflection into selective, audience-aware final messages.
1 Motivation and Objective
Stance preparation turns an initial reaction into a discussable claim while preserving user control over what reasoning becomes visible. StanceLab investigates conversational AI support for clarification, organization, challenge, and audience-aware writing.
- The process includes clarifying beliefs, organizing ideas, identifying weak assumptions, and anticipating challenges from others.
- Stance preparation turns an initial reaction into a claim that can be discussed, revised, or selectively shared.
- Writing decisions include preserving personal voice, packaging reasoning for an audience, and deciding what should remain private.
- StanceLab asks how conversational AI can support pre-deliberation while preserving the user's ownership of the final post.
2 Related Work
The paper combines reflection support, opinion exploration, and conversational-agent role design for preparing a personal stance before deciding whether and how to join a discussion.
- Prior reflection systems show that conversational prompts can scaffold reflection in everyday and workplace contexts.
- Opinion exploration and deliberation research highlights the value of structured exposure to multiple perspectives.
- Conversational-agent research suggests that an agent's role and communication style can change the user experience.
- StanceLab applies these ideas to preparing a personal stance before a user decides whether and how to join a discussion.
3 Prototype
StanceLab is a prototype that separates persona-based diagnosis from user-controlled writing. It supports parallel role responses and a workspace for turning selected insights into a short written stance.
- StanceLab uses a preparation workflow in which an initial reaction enters reflection, persona diagnosis, and message writing.
- The three-persona mode sends each user message to Interviewer, Mentor, and Opponent prompts in parallel with shared context and prior role responses.
- Users can inspect responses, continue conversations, reply to roles, and move selected material into a draft area.
- Both modes begin with the intended sharing context, initial opinion, confidence level, and issue to explore.
- The writing workspace lets users turn selected diagnoses into a documented stance oriented toward a possible audience.
4 Method
The formative within-subject pilot compared three-persona and standalone reflection across 12 sessions, using task outputs, logs, interviews, and observations. The study produced descriptive evidence while varying topics and task framing across sessions.
- Six participants completed two stance-reflection tasks using both the three-persona and standalone modes.
- The tasks covered a shared community-scale prompt and participant-chosen personal issues such as study, career, organizational, and AI-development decisions.
- Mode order was counterbalanced, but topic sensitivity and task framing varied across sessions, making the mode comparison descriptive.
- Each session included an initial stance, a 0–100 confidence rating, system conversation, and a short final notepad message.
- The analysis combined JSON-log and notepad summaries with reviews of transcripts, drafts, interview materials, and observation notes.
5 Current Results
The pilot found that multi-persona support was most useful when roles produced actionable diagnoses, but separate parallel responses created coordination work. Participants also wanted reflection selectively integrated into writing while retaining control over public disclosure.
- Task completion and engagement: All 12 sessions ended with a short final message, while 146 user turns indicated engagement beyond one-shot prompting.The standalone mode averaged more turns than the three-persona mode, and confidence data were incomplete and topic-sensitive.
- Persona diagnosis: Persona responses helped when they surfaced a specific question, objection, weak assumption, or useful reframe that changed what participants could do next.Participants valued challenges that exposed an undefined problem or converted scattered thoughts into a draftable structure.
- Parallel-response coordination: Separate persona responses created coordination load, with participants describing the interaction as managing several simultaneous conversations or feeling publicly judged.Participants often wanted the system to integrate responses before asking them to reply.
- Persona role differentiation: Persona value varied by role: the Mentor organized context and suggested improvements, while the Opponent was useful when relevant but wasteful when off-topic.The Interviewer was valued for specific clarification questions but could overlap with the Mentor or become vague agreement.
- From reflection to writing: Participants wanted a writing workflow that diagnoses problems first, then supports integrated revision and selective transfer into a draft.Some participants copied useful points and revised tone themselves, while others preferred an integrated LLM to address identified problems sequentially.
- Selective disclosure: Sharing remained selective: some participants treated outputs as private reasoning, whereas concrete implementation details could make a message feel suitable for sending.Participants distinguished private preparation from public expression and considered audience-facing reasoning and actionability when deciding whether to share.
6 Limitations and Future Research Plan
The formative pilot is treated as design evidence rather than evidence for general deliberation-quality claims. Future work will evaluate sharing, recipient responses, and downstream discussion while developing an integrated diagnosis-and-writing workflow.
- Limitations: The small formative pilot supports design exploration, not general deliberation-quality claims.Participants came from a nearby student developer community, and the comparison used different topics with varied sensitivity and audience.
- Future research: Future studies should measure whether participants share final messages, how recipients respond, and how prepared stances shape downstream discussion.These studies should preserve the privacy boundary that makes preparation useful.
- Future research: The next StanceLab version will integrate persona diagnosis with writing by identifying key stance needs, synthesizing them into a diagnosis card, and supporting a writing workspace.The workflow targets the strongest claim, missing context, likely audience misunderstanding, and useful objections.
7 Conclusion
StanceLab supports stance preparation when persona roles surface useful blind spots, questions, or challenges. The conclusion favors a diagnosis-and-writing workflow that helps users selectively turn those insights into personally controlled, audience-appropriate stances.
- Conclusion: Multi-persona AI can support stance preparation when persona roles surface a useful blind spot, question, or challenge.This conclusion identifies the usefulness of the roles' diagnostic contributions as the condition for support.
- Conclusion: Parallel persona responses create work when they remain separate from drafting.The conclusion links this coordination problem to the separation between reflection and writing.
- Conclusion: A diagnosis-and-writing workflow should use persona diversity to identify what needs attention before users selectively create personally controlled, audience-appropriate stances.The proposed direction connects persona-based diagnosis with selective drafting.