Developer Antimatter introduces a wordplay game that challenges players to match opposite word tiles using a Mahjong solitaire mechanic. The creator spent extensive time playtesting various mechanics before settling on opposites, which they found work particularly well with the familiar tile-matching format. The game represents an innovative approach to word puzzles that leverages semantic relationships.

Building the puzzles proved challenging, as current LLMs struggle to generate more complex word associations beyond basic opposites like hot-vs-big-vs-fast. The developer notes that new inferences tend to circle a small pool of concepts unless prompting can guide the model into new linguistic territories. This limitation led to the need for graph traversals in puzzle generation.

The game features 20 algorithmically generated levels built from a massive semantic graph containing over 100M edges. This graph combines manual lexicography with millions of LLM inferences from various models. Players can reload to see different randomly selected puzzles, showcasing the system's ability to generate diverse challenges from the same underlying data structure.