I’m a daily New York Times crossword player and I’ve built xword-pipeline with the goal of creating my own dense New York Times style crossword puzzles.

The pipeline consists of a Rust grid fill-engine and a Claude based clue-writer. The fill-engine generates candidate grid skeletons (default is 200) and fills them with words from the Crossword Nexus Collaborative Word List. The clue-writer uses Claude to generate multi-tiered clues (easy, medium, hard), explanations, and QA review the generated puzzles. The clue-writer explanation and/or QA steps can be skipped or the clue-writer can be skipped entirely if you want to write your own clues.

I’ve also created a crossword puzzle web app that can be used to test puzzles generated by the pipeline (https://wordfuzz.com/). The puzzles in the archive have all been generated with the pipeline. QA reviews did show issues with some of these puzzles but they should be solvable and I’ve left them as is so the quality of the pipeline can be accurately judged. You can run the pipeline yourself and use the test page ( https://wordfuzz.com/test) to test the puzzles you generate. The test page uses LocalStorage so nothing is uploaded to the server.

Note: The clue-writer requires an Anthropic API key and costs me an average of around $0.50 to fill a 15x15 grid with multiple-tiered clues and explanations. If you’d like to give the pipeline a try, run a single puzzle first and verify the cost is acceptable to you before generating multiple puzzles.

Any feedback on the pipeline would be appreciated. I would be great to hear if you’ve had success generating fun, playable crosswords or if you have any suggestions for improving the pipeline. Thanks!