Constella: Supporting Storywriters' Interconnected Character Creation through LLM-based Multi-Agents

Creating a cast of characters by attending to their relational dynamics is a critical aspect of most long-form storywriting. However, our formative study (N=14) reveals that writers struggle to envision new characters that could influence existing ones, balance similarities and differences among characters, and intricately flesh out their relationships. Based on these observations, we designed Constella, an LLM-based multi-agent tool that supports storywriters’ interconnected character creation process. Constella suggests related characters (FRIENDS DISCOVERY feature), reveals the inner mindscapes of several characters simultaneously (JOURNALS feature), and manifests relationships through inter-character responses (COMMENTS feature). Our 7-8 day deployment study with storywriters (N=11) shows that Constella enabled the creation of expansive communities composed of related characters, facilitated the comparison of characters’ thoughts and emotions, and deepened writers’ understanding of character relationships. We conclude by discussing how multi-agent interactions can help distribute writers’ attention and effort across the character cast.

💡 Research Summary

Constella is a novel creative‑support tool that leverages large‑language‑model based multi‑agents (LLM‑MA) to help fiction writers design casts of inter‑related characters. The authors begin by conducting a formative interview study with fourteen writers, uncovering three core pain points: (1) difficulty imagining new characters that can meaningfully influence existing ones, (2) challenges balancing similarity and difference across a character ensemble, and (3) the need to flesh out relational dynamics through explicit interaction scenes. From these observations they derive three design goals: suggest diverse, relationally linked characters; enable deep comparison of characters’ inner worlds; and make relationships concrete through simulated exchanges.

Constella implements three social‑media‑inspired interaction modules, each powered by a GPT‑4 backend and orchestrated as independent agents that exchange prompts in a round‑based simulation.

• Friends Discovery takes a seed character profile and generates three mini‑profiles that are relationally tied (e.g., friend, rival, mentor). Each generated “friend” is produced by a distinct agent with its own objective, ensuring variety while preserving a coherent link to the seed.

• Journals produces diary entries for a set of characters around a shared theme (e.g., “first battle”). The agents condition their output on each character’s established traits, back‑story, and emotional state, yielding differentiated inner monologues that writers can juxtapose to surface subtle contrasts.

• Comments lets characters respond to one another’s journal entries, effectively simulating a comment thread. By feeding the prior journal text back into the agents, the system generates replies that reflect the responder’s relationship type and affective tone, surfacing hidden tensions or alliances.

Technically, the system uses pre‑defined prompt templates for each module, and the multi‑agent loop iteratively updates the shared context so that later rounds respect earlier generated content. This design keeps the writer in control: the AI’s suggestions are presented as scaffolding material (profiles, journal excerpts, comments) rather than ready‑to‑publish prose, preserving authorial agency.

To evaluate Constella, the authors ran a 7‑8‑day deployment study with eleven professional and hobbyist writers tasked with building a full‑length story. Participants used the tool to create character pools, back‑stories, scene outlines, and overall plot structures. Qualitative interviews and interaction logs reveal three major outcomes. First, the Friends Discovery feature enabled writers to surface peripheral characters they would otherwise overlook, expanding the narrative world. Second, the Journals interface facilitated systematic comparison of characters’ motivations and emotional responses, often sparking new scene ideas. Third, the Comments feature materialized relational dynamics, making latent conflicts or camaraderies explicit and giving writers concrete hooks for plot development. Importantly, participants reported that the AI’s output felt advisory rather than prescriptive; they retained full editorial control and felt their creative voice remained dominant.

The paper’s contributions are threefold: (1) the introduction of Constella, the first LLM‑MA system explicitly designed for relational character creation; (2) empirical evidence from a real‑world deployment that multi‑agent interactions can offload cognitive effort and promote comparative thinking among writers; and (3) a design argument for AI‑assisted creative tools that prioritize scaffolding over direct text generation, thereby safeguarding authorial agency. The authors suggest future work on finer‑grained agent control mechanisms, integration with long‑term story‑arc planning, and broader evaluation across genres and cultural contexts.