Dartwords swaps the puzzle grid for a chatty assistant, and inherits the blame for bad clues

The Game Band, the studio behind Blaseball, launched a daily word game called Dartwords on the web today ^[1][2]. It starts you with no hints at all: you type any word, and a Clippy-inspired character named Darty answers with a clue about the day's secret word, across up to ten guesses ^[3][4][5].

The clue can be a category or a comparison. Guess a type of animal and Darty might say the answer is bigger, or that it is manmade ^[6]. Progress is shown as a dart landing closer to a bullseye as you get warmer ^[7]. Because nothing is revealed before the first guess, every piece of information in a round comes from Darty reacting to words the player chose ^[8].

Underneath, creative director Sam Rosenthal says the studio built a bespoke hint engine using a local LLM trained on the English dictionary, with broader clues generated by the model and clues written by the developers taking over as you close in on the answer ^[9][10]. He calls it "this interesting mixture of using hand-authored work and local models to get us to where we want to be" ^[10]. Rosenthal says the art and UI/UX are handmade, and that the team uses tools like Claude Code for day-to-day coding "but we pair them with our taste, judgment, and careful reviews" ^[11].

The design consequence is variance. Fixed feedback can be unhelpful but it is never wrong; a generated hint can be wrong, and the player knows it instantly. Rosenthal has priced that in. Darty is a goofy sentient dart with big expressive eyes and tiny hands for clapping ^[12], and the reason given is blunt: "we need something for people to get mad at if it gives you a bad clue" ^[13]. That is a support strategy as much as an art direction.

The second consequence is that the quality curve runs backwards. Rosenthal says that with more people playing, the system will get better and hopefully serve better clues ^[14]. Launch week is when the engine knows least and the audience is largest.

The Verge's reviewer reports it took a few rounds to get the hang of, and that the open-ended format can be "almost paralyzing" because you have to think of a word likely to produce a useful hint, before eventually solving one in three guesses ^[15]. That is the real cost of dropping deterministic feedback: the player has to invent the probe, not just read the result.

The field is busy. Rosenthal's stated goal was something "more conversational" than its contemporaries ^[16], which is a reasonable place to look for room given that Wordle's success produced a deluge of daily word games, with LinkedIn and The Atlantic launching puzzles and platforms like Puzzmo running their own ^[17]. The studio's own line is eclectic rather than focused: Where Cards Fall, then Blaseball, then Dead Man's Party, a Knives Out-themed title in Netflix's push into TV games ^[18].

Worth watching: whether clue quality visibly improves as Rosenthal predicts ^[14], whether hand-authored endgame clues can be produced at a daily cadence ^[10], and whether an open guess space yields anything as compact to share as the score Rosenthal himself passes around to friends and family ^[19]. The source material does not describe how Dartwords handles sharing, which is the mechanic that built this category.