StarCharM: GenAI NPC Creator for Stardew Valley
- StarCharM is a GenAI-based system that transforms free-form character descriptions into structured, game-ready NPC mod files.
- It employs a multi-stage prompt chaining workflow with schema-constrained JSON to generate detailed profiles, dialogues, and schedules.
- User studies indicate that the tool lowers modding barriers and boosts creativity, while also raising concerns about originality and narrative depth.
StarCharM is a GenAI-based non-player character creator for Stardew Valley designed to let players iteratively create new NPC mods with minimal user input while preserving fine-grained adjustment through user control (Miralvand et al., 18 Jul 2025). It was introduced as a case study in democratizing game modding with GenAI, with a particular emphasis on NPC creation, and was evaluated through a user study with ten Stardew Valley players who had varied mod-usage experiences but no prior mod-creation experience (Miralvand et al., 18 Jul 2025). The system combines prompt chaining, schema-constrained JSON generation, editable intermediate representations, and export to Stardew Valley’s Content Patcher format, while the accompanying study examines both the enabling effects of such tooling and concerns about originality, trust, and community engagement.
1. Research Context and Design Objective
The motivating premise behind StarCharM is that game modding offers unique and personalized gaming experiences, but the technical complexity of creating mods often limits participation to skilled users (Miralvand et al., 18 Jul 2025). The project therefore explores a future in which every player can create personalized mods for their games, using Stardew Valley NPC creation as a concrete design space.
Within that scope, StarCharM is not framed as a generic mod generator. It is specifically an NPC creator for Stardew Valley, and its central design goal is iterative creation rather than one-shot synthesis. The system accepts a free-form description of a desired character and progressively transforms that description into structured character traits, dialogues, schedules, and final configuration files. This suggests a mixed-initiative orientation: the model automates file construction and content expansion, while the player retains explicit opportunities to steer, revise, and replace outputs at multiple points in the workflow.
The study around StarCharM also positions the system as an intervention into co-creative modding practices. Participants were asked not only whether the tool was usable, but also how it affected ideation, authorship, craftsmanship, and the social dimension of modding (Miralvand et al., 18 Jul 2025). As a result, StarCharM functions simultaneously as a software artifact and as an empirical probe into how GenAI may reshape mod creation.
2. Generative Architecture and Formal Workflow
StarCharM uses OpenAI’s GPT-4o via the OpenAI API and organizes generation through a prompt-based workflow described as prompt chaining (Miralvand et al., 18 Jul 2025). The pipeline decomposes NPC creation into three stages, each producing a more constrained and game-relevant representation than the previous one.
| Stage | Input | Output |
|---|---|---|
| Character Highlight Generation | User’s free-form description | Three highlight cards |
| Detailed Trait Expansion | Selected highlight JSON + original description | Personality, sample dialogues, summary routine |
| Configuration File Generation | Generated JSON | manifest.json, content.json, dialogues.json, schedules.json |
In Stage 1, few-shot prompts produce three highlight cards from the user’s free-form description. Each highlight contains a name, age, birthday, title, four bullet-point traits, and a quote (Miralvand et al., 18 Jul 2025). In Stage 2, the selected highlight, represented as JSON, is combined with the original description, and GPT-4o generates a more detailed profile: characteristics, job, hobbies, food preferences, manners, social anxiety, optimism, sample dialogues, and a summary routine. In Stage 3, GPT-4o formats the generated JSON into Stardew Valley–compatible files: manifest.json, content.json, dialogues.json, and schedules.json.
The data interface is intentionally constrained. User input is free-form text of at least 50 characters describing a desired NPC, including personality, backstory, and hobbies (Miralvand et al., 18 Jul 2025). Every LLM response is validated against a JSON schema; mis-formatted outputs are auto-reformatted or retried up to five times. Prompt engineering is maintained with ChainForge, using clear instructions, few-shot examples, and explicit “avoid reusing examples” clauses. The three JSON stages are then merged into mod files, including metadata in manifest.json, traits and file pointers in content.json, time/location keys and giftDialogues in dialogues.json, and day-of-week strings such as 900 Saloon 39 18 2 /… in schedules.json (Miralvand et al., 18 Jul 2025).
StarCharM also includes gift preference matching. It extracts keywords from personality descriptions and matches them against in-game items using OpenAI’s text-embedding-ada-002. The matching uses cosine similarity,
to pick the top-3 items per keyword for “love,” “like,” “dislike,” and “hate” (Miralvand et al., 18 Jul 2025). This part of the system is notable because it ties free-form semantic characterization to a bounded in-game item ontology without requiring the player to enumerate gift preferences manually.
3. Interaction Design and Mechanisms of User Control
The interface is organized as an iterative wizard-style workflow spanning four pages (Miralvand et al., 18 Jul 2025). Page 1 accepts a free-form description and exposes a “Create Character” action that leads to three generated highlights. Page 2 presents those highlights as cards containing a portrait, name, title, four bullet points, and a sample quote, along with Pin, Regenerate, and View controls. Page 3 is a detailed view divided into four sections: Basic Info, Personality, Dialogues, and Simplified Schedule. Page 4 presents the final configuration, including full schedules.json, dialogues.json, giftDialogues, and a Download Files action.
The control structure is central to the system’s design. Regenerate lets players refresh individual highlights to explore diversity, while Pin locks preferred highlights so users can mix and match generated alternatives (Miralvand et al., 18 Jul 2025). After selecting a highlight, users can modify the expanded character through inline Edit buttons for name, age, birthday, and gender; dropdown menus for manners, social anxiety, and optimism; and free editing of dialogues and schedule summaries. Navigation includes Back, which returns one step, and Back to Main, which restarts the process with a new description.
This interaction model is significant because StarCharM does not treat LLM output as final. Instead, it exposes several intermediate representations that are legible to players and editable before export. A plausible implication is that the design tries to reconcile two competing requirements identified elsewhere in the paper: lowering the barrier to entry and preserving user ownership. The final Download button exports a .zip or folder of all configuration files ready for Stardew Valley’s Content Patcher (Miralvand et al., 18 Jul 2025), so the workflow terminates in a directly installable mod artifact rather than in a textual specification.
4. Empirical Evaluation
The evaluation involved Stardew Valley players, evenly split by reported gender as 5 F and 5 M, aged 18–45 (Miralvand et al., 18 Jul 2025). Their playtime ranged from 20 to 2000 hours. Mod-usage experience was distributed across None (4), Low (3), and High (3), and none had prior mod-creation experience.
The study design had three phases. The introduction session lasted 70–90 minutes and included a pre-study survey, a demo of StarCharM, think-aloud NPC creation, guided installation, and collection of initial impressions (Miralvand et al., 18 Jul 2025). This was followed by an independent period of 6–15 days during which participants freely created and installed NPCs, explored gameplay with mods, and took notes. The study concluded with an exit interview of approximately 60 minutes using semi-structured questions on usability, creativity, challenges, and community impact.
The analysis used inductive thematic analysis of transcripts following Braun and Clarke (2006), with coding followed by affinity diagramming to derive four high-level themes: Creative Process, Positive Perceptions, Challenges, and Community Effects (Miralvand et al., 18 Jul 2025). For an encyclopedia treatment, the methodological significance lies less in statistical generalization than in the structured identification of tensions surrounding GenAI-assisted modding. The paper does not claim large-sample representativeness; rather, it uses a focused qualitative study to surface design implications and value conflicts.
5. Reported Benefits, Frictions, and Failure Modes
Participants reported strong positive effects on ideation and perceived accessibility. Under the theme of democratization and ideation, one participant described the process as something that “sparked my creativity… felt like a writing exercise,” while novices characterized such tools as especially appealing (Miralvand et al., 18 Jul 2025). The tool also supported ownership and pride: participants described satisfaction in seeing their own concepts materialize in the game world, and even those who supplied minimal input still regarded the resulting NPC as their own idea.
A second benefit concerned workflow familiarity. The paper reports that the chatbot-style UI eased adoption and that iterative refinement felt natural (Miralvand et al., 18 Jul 2025). This is important because the interface did not require prior knowledge of mod file schemas or scripting conventions. In that sense, StarCharM reduced friction not only in content generation but also in the transition from ideation to a deployable mod.
The study also identified several failure modes. First, complex visions remained difficult to realize. The AI struggled with nuanced humor and deep backstories; one example was a father character whose “tongue-in-cheek” tone was missed (Miralvand et al., 18 Jul 2025). Second, inaccuracies affected trust. Participants encountered misplaced schedules, such as mismatches between intended and generated locations, and out-of-place dialogues that required manual fixes. Third, content staleness reduced long-term engagement. Repetitive or “robotic” lines led some participants to describe the NPCs as feeling “kind of dead” after repeated interaction (Miralvand et al., 18 Jul 2025).
These findings qualify any simple narrative of automation as straightforward improvement. The paper presents StarCharM as effective for lowering entry barriers and stimulating creativity, but it also documents that rich characterization, tonal specificity, and world-consistent behavior remain difficult to sustain through generation alone. A plausible implication is that GenAI can scaffold early-stage authoring more reliably than it can fully replace careful narrative design.
6. Community Effects, Design Guidelines, and Open Tensions
At the community level, the paper reports an explicit tension between accessibility and craftsmanship (Miralvand et al., 18 Jul 2025). Newcomers felt empowered by the lowered barrier to entry, while some more experienced participants worried that effort previously associated with mod creation could be devalued if “everyone can do it by asking AI.” The study also raised concerns about reduced peer collaboration if AI systems displace forums and tutorials as the primary route into modding. Participants described the possibility of losing “the community element of constructing something together,” and they raised ethical and originality concerns around authorship, transparency, and mod-sharing norms.
The design recommendations follow directly from these observations. The paper proposes adaptive interfaces that begin with simple free-form input and expose advanced modules such as a dialogue editor, schedule editor, and trait sliders for more experienced users (Miralvand et al., 18 Jul 2025). It also recommends a modular workflow in which users can jump directly to any stage, including skipping highlights to move straight to full configuration, or switch between auto and manual modes. Real-time previews are proposed so that changing a single trait propagates immediately to dialogues and schedules. Knowledge graph integration is suggested to keep NPCs consistent with user-installed mods by querying a live mod metadata graph. Mixed-initiative collaboration is recommended as well, including co-editing of AI outputs, provenance tracking for authorship credit, and embedded community feedback loops.
The paper finally formulates the broader trade-off landscape. Automation lowers barriers but risks homogenization and craft devaluation; fine-grained control fosters originality but raises complexity; and balancing AI empowerment with community norms and skill development is described as critical to sustaining vibrant modding ecosystems (Miralvand et al., 18 Jul 2025). In that sense, StarCharM is best understood not merely as a mod generator, but as an empirical instance of a larger design problem in GenAI-assisted creative tooling: how to make creation easier without erasing the practices, standards, and social relations that made the activity meaningful in the first place.
