Constella: LLM-Based Storytelling Assistant

• Constella is an LLM-based multi-agent tool that generates interconnected character networks using features like FRIENDS DISCOVERY, JOURNALS, and COMMENTS.
• It facilitates rapid character creation with contextual mini-profiles, detailed diary entries, and realistic dyadic dialogue simulations.
• Empirical evaluation showed the tool enhanced ensemble expansion and balanced narrative attention, supporting robust creative processes.

Constella refers to an LLM-based multi-agent tool developed to support storywriters in creating interconnected casts of characters and enriching their relational dynamics (2507.05820). Designed in response to empirical findings from studies of practicing writers, Constella directly addresses key challenges in long-form narrative authorship: the imagination of new, contextually relevant characters; the balancing of similarities and differences amongst characters; and the intricate fleshing out of inter-character relationships. By leveraging LLMs orchestrated as interacting agents, Constella provides structured, interactive affordances that scaffold the creative process of constructing, exploring, and deepening fictional communities.

1. System Overview and Motivation

Constella was motivated by formative studies involving experienced storywriters, which uncovered persistent difficulties in character ensemble construction. Writers reported struggling with envisioning new supporting characters who could productively influence existing ones, maintaining narrative balance in character traits, and articulating nuanced relationships that evolve through the story. Addressing these issues, the research team formulated design goals rooted in facilitating interconnected and differentiated character creation while allowing writers to maintain authorial control.

The resulting system operationalizes these goals through three principal, LLM-enabled features—FRIENDS DISCOVERY, JOURNALS, and COMMENTS—that together encourage the formation of expansive, relationally rich fictional communities.

2. Core Features and Multi-Agent Framework

Constella’s architecture is distinguished by its multi-agent approach, wherein each character is modeled as an agent with individual traits, histories, and interrelations. The primary features are as follows:

1. FRIENDS DISCOVERY
   • Writers provide a prompt specifying a desired relationship to an existing character (e.g., “Binggu’s greatest enemy”).
   • The system generates three succinct character profiles, each with a name, a mini-biographical sketch, a backstory, and explicit, bidirectional relationship descriptions detailing the connections between the new and existing character.
   • This social-media-inspired “mini-profile” approach facilitates the organic expansion of the cast, helping writers rapidly explore narratives where new characters are immediately situated in the social fabric of the story.
2. JOURNALS
   • Writers can invoke this feature to probe the inner thoughts and emotional landscapes of one or more characters in response to a scenario or narrative prompt.
   • The system produces “diary entries” for each selected character, written in a first-person style (“Dear Diary …”) and reflecting individualized reactions, interpretations, and emotive responses.
   • Presenting multiple journal entries side-by-side makes distinct perspectives salient, allowing for nuanced comparison and the emergence of unique intra-group dynamics.
3. COMMENTS
   • This feature enables characters to respond to others’ journal entries, creating threaded, conversational exchanges.
   • The format is dyadic and open-ended; for instance, after reading the journal entry of one character, another character can initiate a reply and subsequent back-and-forth can ensue.
   • Through these exchanges, writers can explore relationship development and latent tensions, and simulate plausible dialogue or commentary between members of the ensemble.

These features collectively instantiate a multi-agent simulation in which characters are not isolated templates but active participants in dynamic relational networks. The system iteratively elicits and displays the cast’s evolving social structure, inner life, and dialogue, adhering closely to the principles articulated by the authors.

3. Empirical Evaluation and User Study

Constella underwent a 7–8 day deployment paper with 11 storywriters, with participants engaging in tasks such as character generation, backstory development, scene writing, and outlining. Usage logs, diary studies, follow-up interviews, and a Creativity Support Index (CSI) survey comprised the data collection protocol.

Findings from this paper demonstrated that:

• FRIENDS DISCOVERY enabled writers to generate an average of 3.73 new, contextually coherent characters per author, supporting the growth of richly interconnected ensembles.
• JOURNALS was the most frequently used feature, yielding nearly 539 diary entries; participants noted that it facilitated immersion in the diverse mindsets of multiple characters and supported comparative reflection.
• COMMENTS, while used less frequently, proved valuable for focused exploration of dyadic relationship nuances.
• Collectively, these tools led to the creation of more expansive communities, facilitated comparative analysis of thoughts and emotions, and encouraged more balanced creative attention across the cast—without diminishing writers’ sense of ownership or control.

No formal mathematical notation is introduced, as the system’s contributions are qualitative and behavioral, corroborated by interaction logs and participant narratives rather than formulaic models.

4. Design Principles and Interaction Paradigm

The tool’s design draws upon social media conventions—mini profiles, personal diary entries, and comment threads—not as superficial metaphors, but as functional templates that familiarise and streamline the creative exploration of character relationships. By leveraging these interaction paradigms, Constella lowers cognitive barriers to the extension and deepening of fictional worlds, scaffolding both the discovery of new narrative opportunities and the subtle elaboration of inter-character dynamics.

Furthermore, the explicit focus on bidirectional and threaded relationship modeling enables writers to trace the reciprocal, evolving nature of fictional relationships, supporting more substantive, non-stereotypical character development.

5. Implications for Creative Writing Practice

Constella’s multi-agent architecture and feature set demonstrate that LLM-based tools can move beyond recommendation or “one-shot” text generation, operating instead as collaborative partners in complex, socially-embedded creative processes. The system successfully addresses persistent challenges in narrative production—character differentiation, networked relationality, and distributed authorial focus—by foregrounding relationship emergence and reflection across the entire cast.

A plausible implication is that such tools may help redistribute narrative attention, enabling the emergence of richer ensemble structures and reducing over-reliance on single protagonists or dyadic archetypes. The deployment paper also suggests that tools of this kind, by integrating LLMs as multi-agent systems, may set new precedents for interactive creative support, especially in domains requiring the management of numerous, interdependent entities.

6. Limitations and Directions for Future Research

While Constella was shown to be effective in the context of the deployment paper, its current implementation reflects several limitations:

• The COMMENTS feature, while functionally sound, saw less frequent usage, possibly indicating that writers prioritize independent introspection over interactive dialogue in certain workflows. This suggests a need for further refinement or contextual prompts to maximize engagement with relationship simulation.
• No formal assessment was made of long-term impact on full-length manuscript development or publication outcomes.
• The system’s reliance on static prompt templates and synchronous multi-agent interactions may be extended through adaptive or asynchronously evolving agent behaviors in future work.

Recognition of these limitations points towards further research on integrating multi-agent, LLM-powered creativity support with more sophisticated narrative planning and world-building frameworks.

7. Summary Table: Principal Features of Constella

Feature	Functionality	Role in Creation
FRIENDS DISCOVERY	Generates mini profiles of related characters with mutual relations	Ensemble expansion
JOURNALS	Produces diary entries from one or more character viewpoints	Perspective immersion
COMMENTS	Supports threaded, dyadic interactions as character responses	Relationship evolution

Constella exemplifies the application of LLM-based multi-agent systems to the domain of storywriting, directly supporting the construction, reflection, and expansion of interconnected narrative character networks. Its deployment validates the utility of such tools in augmenting both the breadth and depth of character-driven storytelling within writer-centric creative processes.