Design Techniques for LLM-Powered Interactive Storytelling: A Case Study of the Dramamancer System

AI summary

Overview

Research area: Human-Computer Interaction (cs.HC), specifically interactive narrative and LLM-powered storytelling systems. The work is affiliated with Midjourney (authors include Tiffany Wang, Yuqian Sun, Yi Wang, Melissa Roemmele, John Joon Young Chung, and Max Kreminski; arXiv:2601.18785v1, published 26 Jan 2026, licensed CC BY-NC-SA 4.0).

Technical level: Intermediate. The paper assumes familiarity with interactive narrative concepts such as storylets and conditions/outcomes, and with the idea of using LLM prompting to generate structured narrative content.

Scope: A design case study of the Dramamancer system that catalogs its interface design, its two-module LLM architecture, and the variables relevant to evaluating the playthroughs it produces.

What This Paper Is About

Interactive narrative requires authors to anticipate the consequences of many possible player choices, which is a long-standing authorial burden. LLMs are often proposed as a way to remove that burden by generating story text that accommodates player agency within author-defined boundaries, but it remains unclear what the author, the player, and the LLM should each contribute. The paper uses Dramamancer — a system that transforms author-created story schemas into player-driven playthroughs via an LLM — as a concrete case to surface design techniques and open evaluation questions.

Key Contributions

1. A two-role interface specification. The paper details Dramamancer's author interface (where a story schema is built from style, characters, scenes, and events) and its player interface (where a playthrough unfolds line by line and the player periodically supplies input in the format (actions) dialogue).
2. A two-module LLM system design. It describes an instantiation module that produces the next line of the playthrough and an interpretation module that determines which event conditions the current playthrough satisfies, each implemented as a single LLM interaction.
3. A prompt-design account. It explains specific expectations encoded in the instantiation prompt (a boolean pause variable, restricting generated lines to non-player characters, incorporating satisfied event outcomes organically, and high responsiveness to the player's most recent input) and in the interpretation prompt (returning a list of satisfied conditions).
4. A two-perspective evaluation framework. It proposes author-side variables (style adherence, character distinctiveness, scene awareness, event detection accuracy, outcome realization) and player-side variables (responsiveness, timing, reflection, engagement) as the dimensions along which playthrough quality should be judged.

Main Findings

• Division of labor among author, player, and LLM: The author establishes a story schema; the LLM dynamically realizes that schema as a concrete story instance; the player's input drives the player character's actions and dialogue.
• Story schemas are structured around storylets: Events are modeled as storylets, each pairing a condition (a true/false statement evaluated against playthrough state, which may be null and replaced by a trigger after a set number of generated lines) with an outcome (a description of what should happen, which may end the scene and transition to another scene).
• Pause is an explicit model output: The instantiation prompt asks the LLM to return a boolean pause value, expected to be true when there has been meaningful progression since the player last contributed input.
• Player-character content is reserved for the player: Generated lines are expected to pertain only to non-player characters, so the player character's actions and dialogue come only from player input.
• Triggered outcomes are injected into the prompt: Outcomes of satisfied events are listed and the LLM is instructed to weave them in organically, allowing an outcome to be conveyed across multiple lines.
• Evaluation is framed as two-sided but not empirically executed: The paper enumerates evaluation variables rather than reporting measurements. No benchmarks, dataset sizes, user-study results, or quantitative metrics appear in the content provided.

Methodology in Plain English

The researchers did not run an experiment. They took an existing system, Dramamancer (previously described in Sun et al., 2025 and Wang et al., 2025), and analyzed it as a design case. They decomposed the system into its user-facing parts (what authors author, what players do) and its technical parts (the instantiation and interpretation LLM calls), then described the design decisions embedded in each. From that decomposition they derived a set of variables that would need to be assessed to judge whether a generated playthrough is good, split according to whether the author or the player is the one judging. The result is a conceptual framing plus a checklist of design and evaluation considerations, not a measured result.

Why This Matters

The paper reframes a common assumption in LLM storytelling research — that LLMs will simply "eliminate authorial burden" — into a concrete question about how responsibilities should be divided among the author, the player, and the model. By naming the specific design choices (pause control, non-player-character-only generation, outcome injection, condition detection) and the specific quality dimensions those choices affect, it gives researchers and practitioners shared vocabulary for comparing LLM narrative systems.

Real-world applications:

• Interactive fiction and narrative games that need to honor authored plot beats while still responding to player choices.
• AI-assisted authoring tools for writers who want to define structure and tone rather than write every branch.
• Tabletop and role-playing game tooling, where a schema of scenes and events could be instantiated live by a model acting as a facilitator.
• Training and educational simulations, where scenario designers need guaranteed coverage of specific events while learners act freely.

Industry relevance is direct: the authors are affiliated with Midjourney, and the case study reads as an internal design rationale for a deployed-style creative AI system, making it useful to teams building LLM narrative products.

Future Directions

• Operationalizing the proposed evaluation variables. Style adherence, character distinctiveness, scene awareness, event detection accuracy, outcome realization, responsiveness, timing, reflection, and engagement are named but not measured here; turning them into testable measures is an open task.
• Determining how much authorial burden actually shifts. Whether schemas genuinely reduce the burden of anticipating player choices, compared to traditional branching authoring, remains unverified in this content.
• Revisiting the division of labor as LLM capabilities evolve. The paper explicitly frames the author/player/LLM split as an open question, implying the answer may change as models improve.
• Generalizing beyond Dramamancer. Whether these techniques and evaluation dimensions transfer to other LLM-powered interactive narrative systems is not established.

Target Audience

HCI and interactive narrative researchers; game writers, narrative designers, and creative technologists building branching or generative story systems; LLM application developers who need structured control over generated narrative content; and students studying the intersection of AI and storytelling. Readers looking for empirical results, benchmarks, or user studies will not find them here — this paper is a design and framing contribution.

Authors’ abstract

The rise of Large Language Models (LLMs) has enabled a new paradigm for bridging authorial intent and player agency in interactive narrative. We consider this paradigm through the example of Dramamancer, a system that uses an LLM to transform author-created story schemas into player-driven playthroughs. This extended abstract outlines some design techniques and evaluation considerations associated with this system.