This is a little daily word game that I ve been working on. I love crosswords and Wordle but they all use the spelling of words as the main constraint. With LLMs we can build Squishy Systems , systems that don t have to be fully procedurally defined. That can give us games that feel more like playing a board game with friends. Games that require a human judge. Because they require a judge they aren t well explored for video games, especially outside of multiplayer.Squishword is a simple first exploration of that. There s a tutorial so I won t explain all the mechanics. Simply: you try to move from one word to another by adding, subtracting, joining and dividing words. For example: dog splits into bark and animal . You can do clever things like take bark + plant and make tree . The inspiration comes from the Wikipedia game where you try to move from one random article to another just by clicking links.Tuning the judge behavior for fair results took a lot of work, and I m not 100% happy with it yet. I built more tools and visualizations than features in the game itself. Like the 11x11 matrix of logic and wit grades for every move, to try to get the judge distribution more even. This is a simple example of LLMs in a product, so I m curious how more complex workflows are dealing with it.