I gave a recommendation around LLM crossword generators a while ago, so reposting it:
As someone who has dabbled in AI generated crosswords I found that providing samples of "good crossword clues" (which I curated from historical NYT monday puzzles) as part of the LLM context helped tremendously in generating better clues.
Part of the deep satisfaction in solving a crossword puzzle is the specificity of the answer. It's far more gratifying to answer a question with something like "Hawking" then to answer with "scientist", or answering with "mandelbrot" versus "shape".
So ideally, you want to lean towards "specificity" wherever possible, and use "generics" as filler.
Producing a valid layout is a different situation. Ultimately depends on how strict you are in terms of acceptable layouts. Getting a NYT compatible crossword puzzle is going to be more difficult by virtue of its symmetrical requirements than if you're okay with just laying out sequences of perpendicular intersecting words.
Thanks for the response. I agree that specificity is important.
I provided a few specific clues and answers along with the grid layout, and a theme as context. Ideally this would generate a filled crossword that I could then verify and edit.
Claude would churn, restarting from scratch over and over trying to fill the layout. If I remove the grid layout from the context, I can generate sparse sequences like you mention but these are not very fun to play.
I had no luck generating full-size, NYT-style puzzles, but was able to generate 5x5 minis!
I pre-defined a handful of layouts (minis generally just have corners blacked out), and brute-forced puzzles with all 4 or 5-letter words (depending on the space available, given the blacked-out squares) using a word list from Wordnik. I used an LLM to generate clues.
Thanks. I wasn't aware of XWords. It works well for generating puzzles with sequences of intersecting words. It doesn't seem to produce denser puzzles as you add more clues and answers. It just builds a larger puzzle.
I gave a recommendation around LLM crossword generators a while ago, so reposting it:
As someone who has dabbled in AI generated crosswords I found that providing samples of "good crossword clues" (which I curated from historical NYT monday puzzles) as part of the LLM context helped tremendously in generating better clues.
Part of the deep satisfaction in solving a crossword puzzle is the specificity of the answer. It's far more gratifying to answer a question with something like "Hawking" then to answer with "scientist", or answering with "mandelbrot" versus "shape".
So ideally, you want to lean towards "specificity" wherever possible, and use "generics" as filler.
Producing a valid layout is a different situation. Ultimately depends on how strict you are in terms of acceptable layouts. Getting a NYT compatible crossword puzzle is going to be more difficult by virtue of its symmetrical requirements than if you're okay with just laying out sequences of perpendicular intersecting words.
Thanks for the response. I agree that specificity is important.
I provided a few specific clues and answers along with the grid layout, and a theme as context. Ideally this would generate a filled crossword that I could then verify and edit.
Claude would churn, restarting from scratch over and over trying to fill the layout. If I remove the grid layout from the context, I can generate sparse sequences like you mention but these are not very fun to play.
I had no luck generating full-size, NYT-style puzzles, but was able to generate 5x5 minis!
I pre-defined a handful of layouts (minis generally just have corners blacked out), and brute-forced puzzles with all 4 or 5-letter words (depending on the space available, given the blacked-out squares) using a word list from Wordnik. I used an LLM to generate clues.
Yeah, it seems the general consensus is that LLMs can't generate large newspaper style crosswords.
For now I will just use Phil to build mini puzzles and try using an LLM to help generate clues. Thanks.
I recently used https://www.xwords-generator.de/en for a party and it turned out great. Supports solution reveal as well.
Thanks. I wasn't aware of XWords. It works well for generating puzzles with sequences of intersecting words. It doesn't seem to produce denser puzzles as you add more clues and answers. It just builds a larger puzzle.
I’m pretty sure this is what the pros use. No AI though https://www.crossword-compiler.com/
Thanks. I'll give that a try. I also found https://crosshare.org/ which is open source and has a crossword constructor.