[vunderba]

Nice. I experimented with automated cruciverbalism in the pre-LLM era, and it was a pretty good challenge especially for strict NYT-style crosswords, which maintain aesthetically pleasing symmetry in the puzzle layout.

Looks like we actually exchanged a bit of conversation on this a while back:

https://news.ycombinator.com/item?id=45097346

I think you've done a good job improving on the clue's specificity. As I mentioned in that thread, part of the deep satisfaction of solving a crossword puzzle is the specificity of the answer. It’s far more gratifying to fill in something like “shortwave” than “radio,” or “gallium” instead of just “metal.”

I attempted (over a decade ago at this point) to manage word types using frequency corpuses to help ensure a healthy balance of specific vs filler answers with varying degrees of success...

Small bit of feedback:

On the 9x9 puzzle for Sunday 2026-07-19, one of the three-letter clues was “Cook of Apple fame.” I don’t think this is much of a spoiler, but the answer in the puzzle was WOZ, referring to Steve Wozniak. However I’ve never heard anyone call him a cook. It's made even more confusing because there is a Cook at Apple, aka Tim Cook who took over as CEO after Steve Jobs.

[ekorbia]

Yes, I tried a while back to use LLMs for both the grid fill and the clue writing but was not able to generate dense puzzles. This pipeline uses a Rust fill engine that fills hundreds of candidate grids using a scored word list.

Thanks for the feedback! I will recheck the clues for that puzzle.