A small look inside the MiniWord sausage factory. This whole thing started as an experiment to learn about crossword construction, and has slowly become more about leveraging LLMs to build tools to improve the process now that I find myself needing to crank one out daily.

LLMs are not great at multidimensional reasoning and zero-shotting a perfect grid. But they're really good at pretending to be an editor, and better than me at least at salting clues to taste for difficulty. Haven't started using the feature here to beef up the challenge on the MiniWord, but we've got some interesting options with something like this.

Tempted to drop a difficulty setting in MiniWord (say: Easy, Challenging, Expert) and have corresponding leaderboards.

If you're solving the MiniWord daily, let me know your thoughts. For now it's fun to see the apples-to-apples comparison of who can do it the fastest. But it's clear there's some of you out there that want a challenge, and that's one click away on my side

crossword constructors marked safe from LLM taking their jobs for now

though they make constructors more efficient

Building @warpcast and @farcaster / dad / like to read, cook, ski, code

ok but can an LLM vibe code updates to react-crossword so when i am entering a word where a tile is already filled it skips over that file

u can just dm next time, id rc this for free lol

https://markedsafe.org/safe/LLM%20Taking%20Their%20Jobs

so, crossword folks are safe from the AI and bot invasion, interesting. 
Same should apply to jigsaw puzzle creators, no?