Dartwords: A Daily Word Game Powered by a Local Language Model
The Game Band has launched Dartwords, a daily web-based word guessing game that pairs a Clippy-style mascot with a locally hosted large language model (LLM) to generate contextual hints. Players get 10 chances to guess a secret word with no initial clues. After each guess, a character named Darty responds with hints, such as whether the answer is bigger or manmade. The Verge
Dartwords was created by Sam Rosenthal, founder, CEO, and creative director of The Game Band. The studio's previous titles include the narrative puzzler Where Cards Fall, the fan-driven baseball simulator Blaseball, and the Knives Out-themed party game Dead Man's Party. The Game Band's homepage tags Dartwords as "BRIGHT, PLUCKY, and SYNCOPATED" under the tagline "The word game that responds to you!" The Game Band
The technical architecture behind those responsive hints is where Dartwords gets interesting. The Game Band built a custom hint engine that runs a local LLM — a text-based AI model running on their own infrastructure rather than calling a cloud service — trained on the English dictionary. The system combines hints generated by the model with clues written directly by the developers. That hybrid approach, using a model scoped to a dictionary rather than an open-ended body of text, limits the model's output to a narrow domain where hallucination risk (the tendency of language models to produce confident but false information) is lower and hint quality is more predictable. The Verge
Rosenthal stated that all of Dartwords' art and UI/UX is handmade. For day-to-day coding, the team uses Claude Code, Anthropic's agentic coding tool — a system that can write and edit code based on developer instructions. The split is deliberate: generative AI handles routine implementation work, while the visual and interaction design stays in human hands. The Verge
For a small studio, this is a practical illustration of the current tooling stack. A local model fine-tuned on a bounded dataset, a coding assistant for implementation, and handcrafted creative work are not competing approaches here. They are layered into a single production pipeline, each assigned to the part of the work it handles best.
The decision to run the hint engine locally rather than calling a hosted API is worth examining. Running inference (the process of the model generating a response) on studio infrastructure eliminates per-query latency and cost, which matters for a free, daily, web-native game with potentially high session volume. It also removes the failure mode of a third-party API rate limit or outage taking the game offline. The tradeoff is model size: a local LLM scoped to dictionary hints will be smaller and less capable than a frontier model, but the domain constraint makes that an acceptable, even desirable, limitation.
The Clippy comparison is apt and probably intentional. Microsoft's Office Assistant, for all the mockery it attracted in the late 1990s, established a template for a conversational interface that reacts to user input in real time. Darty occupies the same functional niche: a persistent character that acknowledges each guess and nudges the player forward. The difference is that Clippy's responses were scripted and rule-based, while Darty's hints are model-generated and therefore variable across sessions, even for the same daily puzzle.
The Game Band's portfolio gives some context for why this studio would build a game around a locally hosted model. Blaseball was known for its unpredictable, fan-shaped narrative loops. Where Cards Fall focused on spatial, emotional puzzle design. Dead Man's Party brought social deduction into a branded setting. Dartwords is a different genre entirely, but the throughline is a studio that builds systems designed to respond to player input in ways that feel organic rather than pre-scripted. An LLM-powered hint engine is a natural extension of that design philosophy into a word-game format. The Game Band
The broader context here is that the most notable aspect of Dartwords may not be the game itself but the production model it illustrates. A small team can now ship a polished, daily web game with a custom local ML inference pipeline and an AI-assisted codebase, without the infrastructure overhead that would have been required even two years ago. The creative work, the art and UX, remains human-made. The model is bounded to a specific domain where its output is reliable. The coding assistant accelerates implementation. Each layer is doing what it is genuinely good at, and the seams between them are where the craft lies.
In this author's view, that is a more instructive pattern than any single game launch. The studios that figure out where to draw the line between generative and handmade, and how to bind a model tightly enough to its task that its weaknesses stop mattering, will be the ones that ship interesting things at this scale. Dartwords is a small game, but it is a clear example of that approach working in practice.
