Stillwater
Designed with GPT-5.6 Sol.
Can balancing neighboring tanks make a satisfying daily puzzle?
Play Stillwater →The design experiment
What happens when you ask different language models to design a puzzle game, then make their ideas playable? Wit Orchard is an invitation to find out.
Alon set the direction and reviewed the work. The models proposed mechanics, puzzles, and implementation handoffs. Codex helped check the designs and build them into this website. Corrections and human decisions are part of the result.
This is a design case study, not a controlled model benchmark. The conversations, tools, and amount of iteration differed. Passing a solver check establishes a specific property; it does not establish that a game is fun, original, or better than another model’s game.
Play is free. You do not need to give an email address to try the games or read these notes.
Designed with GPT-5.6 Sol.
Can balancing neighboring tanks make a satisfying daily puzzle?
Play Stillwater →Designed with Claude Fable 5.1.
Can one shape be cut into two matching pieces?
Play Twins →Designed with Claude Opus 5.
Can folding and punching paper become a precise spatial puzzle?
Play Foldglyph →Designed with Gemini Flash 3.8.
Can four careful cuts reveal four words?
Play Snipfall →Designed with GPT-6 Astra.
Can two boats share one breeze and use their wakes to get home?
Play Lantern Leeway →An early Snipfall example described as unique admitted ten valid outcomes under the chosen dictionary. The downloadable sample includes the board, a small checker, the exact dictionary, and every outcome. That finding led to an explicit acceptance test.
For Lantern Leeway, the designer disclosed a shortcut in an early teaching level. Later checks examined the revised collection and tested whether the wake mechanic actually mattered. Those checks address the supplied puzzles, not every possible puzzle.
The useful lesson is practical: define what “valid” means, run the check, retain its output, and record which version it applies to.