Can I Create an AI Agent to Create Crossword Puzzles? Part 1: The Adventure Begins

I, like many people, enjoy running through the NYT puzzle rotation as part of my morning routine. The Mini Crossword, Spelling Bee, Wordle, Connections, and if I’m feeling spunky, Letter Boxed, give me a few minutes of distraction while I drink my coffee and steel myself for a long day of Zoom meetings and arguing about Jira stories.

But also like many people, I’m a little hesitant to whip out the credit card for yet another subscription. I’ve already got a thousand subscriptions to a thousand things (even cars are getting in on the action). So I was upset, distraught even, when the NYT recently locked the Mini behind a paywall.

Now, look. I get it. Making crosswords is hard. Making good crosswords is even harder. I’ve tried it before; just building a workable grid is brutal, and that's before you even get to the clue writing. The level of effort and creativity displayed by creators absolutely deserves fair compensation via NYT’s subscription revenue.

And yet, the duality of man. I’m also a software engineer living in the budding age of AI. I see a task that’s difficult and labor-intensive, and the soulless corporate drone in me can’t help but ask: “Could we automate this with AI?”

So here we go.

Is This Even a Good Idea?

There’s this really old dinosaur movie with a famous line: “Your scientists were so preoccupied with whether they could, they didn’t stop to think if they should.”

And that feels oddly appropriate before we even start down this path. It’s 2025, and the world is brimming with fears, some justified and some unjustified, that AI is coming for our jobs. Creative works like crossword puzzles carry a human touch. Clever puns, twists of phrase, cultural references, subtle jokes, all rely on human experience and intuition. How could an AI replicate a puzzle that just feels right?

If you’ve done enough puzzles, you know why people keep coming back to the NYT rather than the bargain bin collections of 100 puzzles for three bucks. There is something about the clues, the flow of answers, the little tricks that authors use. If we outsource this to a cold, emotionless robot, are we just piling another shovel of dirt onto the grave of human creativity? Are we trading connection for robotic efficiency? I don’t know, man.

But hey, with that moment of self-reflection out of the way, let’s splice some frog genes into a T-Rex and see what happens.

Editor's Note: The author of this article is well aware that most of the scientists died in Jurassic Park.

What Tools Are Out There?

Before we dive down the rabbit hole of side projects, let's do something that should probably be more popular with side projects. Let's look at what’s out there. And who better to start with than the masters themselves, the NYT. They publish popular resources for crossword building, and there are several paid and open-source constructors available.

There are many tools, both browser- and desktop-based, that provide a number of important constructor features:

• Word banks and validations. Word lists are actually a surprisingly difficult part of crossword construction. Having a good word list is critical. Sure, you can use a dictionary, but that just gives you the basics. What about abbreviations like UCLA or AARP? Short phrases like BESTMAN? Names like JOEBIDEN or pop culture phrases like DADBOD? None of these are hard to think of or recognize in isolation, but if you are writing software to help find words that fit, you need to account for all of them and then curate the list carefully.
• Grid templates and autofill. Constructor features let you choose or set grid templates, see options that fit, or even automatically fill parts of the puzzle.
• Format support. Many tools support common formats like .ipuz, allowing you to export or import puzzles into different editors.

So yes, there are tools out there, and even some AI tools from side project developers like me. But at least at the time of writing, there is no big player saying, “Of course we have an AI agent that builds crosswords.”

What Happens If I Ask Nicely?

I mean, what if I just... ask an LLM to make me a puzzle? Can't hurt, right? Let's see. I'm using ChatGPT-5.

Well, that went... okay. Not bad, but not fantastic. Ignoring the fact that I asked for a 5x5 and it actually gave me a 4x5, the clues are terrible. One down is "CATSS" and the clue says it's three letters. Two down is "ARIET," which claims to be three letters and, importantly, isn’t even a real word. So yeah, this puzzle sucks.

Let's try again, but with a bit of prompt engineering.

__ASK__
Generate a 5x5 crossword puzzle and associated clues. 

__CONTEXT__
- The puzzle must be returned in the open-source `.ipuz` format, which supports crosswords and other word puzzles.  
- The `.ipuz` format should include grid, numbering, solution, and clue data.  
- Crossword convention should be respected: symmetric grids when possible, no unchecked letters, and no answers shorter than 3 letters.  
- Clues should follow the style of mainstream daily crosswords: concise, fun, and accessible. They should favor common vocabulary, everyday abbreviations, short phrases, well-known celebrities, or pop culture references.  
- Black squares can be used strategically to shape the puzzle and enforce the 3+ letter word length rule.  
- The goal is to create a puzzle that feels natural, solvable, and engaging, rather than overly technical or obscure.  

__CONSTRAINTS__
- All answers must be at least 3 letters long.  
- Use black squares as needed to enforce standard crossword structure.  
- The puzzle should be solvable, cleanly constructed, and symmetric if possible.  
- All answers and clues should come from common words, abbreviations, short phrases, celebrity or common names, or widely recognized pop culture references.  
- Keep clues fun, clear, and accessible—avoid obscure or overly technical material.  
- Return only valid `.ipuz` JSON.  

__EXAMPLES__
{
  "origin":"crosswordrace.com",
  "version":"http://ipuz.org/v1",
  "kind":["http://ipuz.org/crossword#1"],
  "author":"crosswordrace",
  "publisher":"crosswordrace.com",
  "title":"CR Puzzle 10",
  "empty":"0",
  "dimensions":{"width":5,"height":5},
  "puzzle":[["#",1,2,3,4],["#",5,0,0,0],[6,0,0,0,0],[7,0,0,0,"#"],[8,0,0,0,"#"]],
  "clues":{
    "Across":[[1,"Papa's partner"],[5,"They make up everything"],[6,"At-home ab exercises"],[7,"Kilometers, grams, millimeters, etc"],[8,"Type of engineering specializing in machines and manufacturing"]],
    "Down":[[1,"The Pine Tree State"],[2,"Room in a house that's full of memories?"],[3,"Place for a grin or a pout"],[4,"Guitar booster"],[6,"Tallied up"]]
  },
  "solution":[["#","M","A","M","A"],["#","A","T","O","M"],["S","I","T","U","P"],["U","N","I","T","#"],["M","E","C","H","#"]]
}

This time, it gives us a pretty reasonable output. Let's throw the output into an .ipuz viewer and see what we're working with.

Wow, that’s pretty good. The puzzle is sensible, the clues make sense, the grid looks fine, and it generated start to finish in 5 minutes and 55 seconds. That’s pretty quick. Did we just one-shot this?

Is all you need to build an AI crossword puzzle just the right prompting and a few examples?

Conclusion

What have we accomplished in this article? Nothing, really, but we have learned a few things. We have reflected on whether using AI for creative pursuits is even a good idea or if it is totally "missing the point" in the name of efficiency. We have also seen that with a bit of prompt engineering, modern LLMs like ChatGPT-5 were able to successfully generate a reasonable 5x5 mini puzzle.

In Part 2, we will get to building. Let's see how far we can push an AI agent to create something truly playable.