Linking Automation For The NYT Crossword API
I spent about six months building a local script that pulls NYT crossword data, solves clues against an external dictionary, and links results back to the grid automatically. It's not a clean project and it breaks whenever the NYT changes their schema, which is more often than you'd think. The core problem most people hit is that the NYT doesn't offer a public puzzle API anymore. They used to have one under games.nytimes.com, then they removed the access keys. So you're working with either web scraping or the old endpoints that still respond to direct requests if you structure the headers correctly. I ended up going with a headless Chromium instance that loads the puzzle page and intercepts the XHR calls the frontend makes to fetch clue data. The workflow runs like this. You open a puzzle at a set date, scrape the JSON payload from the network request, parse the grid dimensions and theme clues, run each across and down entry against a word list filtered by letter count, then link the best-fit answers back to the corresponding cells. The linking part is where it gets tricky. You need to map the clue index from the NYT response to the actual grid position, which isn't always sequential because of how the official JSON structures the Across and Down arrays separately.
I found that the most reliable approach uses the cell coordinates embedded in each clue object rather than trying to infer position from array order. The NYT response includes a start index and direction, but more importantly it includes a "clueNumber" field that maps directly to grid coordinates when you know the grid size. For a standard 15x15 puzzle, clue number 1 is top-left. Clue number 25 would be somewhere in the bottom row. That mapping breaks down for mini and large puzzles though, so you need separate coordinate calculators per grid type. Here's the part nobody talks about. The NYT grid JSON doesn't always include the answer key. It includes blocker cells marked as black squares, but the solution letters are not embedded in the data you pull. That means the "solving" step has to be entirely external. I built a simple backtracking solver that fills cells using a Scrabble-valid word list, scoring each candidate by letter frequency, bigram probability, and how well it intersects with already-placed answers. It's not perfect. It gets about 78 percent of clues right on first pass for medium-difficulty puzzles and drops to roughly 62 percent on Sunday puzzles because the obscure fill words break the frequency models. The real bottleneck is the linking phase after solving. You take the generated answers and try to overlay them onto the grid. Conflicts happen constantly because two intersecting clues might both resolve to different valid words that share a common letter but differ elsewhere. I solved this by running a constraint satisfaction loop that adjusts both across and down entries simultaneously rather than greedily locking in the highest-scoring word for each clue independently. That usually converges in under 3 seconds for a 15x15 puzzle on a modern laptop.
There's a specific edge case that cost me three weeks to fix. Sometimes the NYT API returns clues in a different order than the printed puzzle, especially on themed Sundays where the theme answers are numbered out of sequence. If you blindly link clue index 3 to grid cell 3, you'll get garbage results. I had to write a heuristic that matches clue text patterns to likely grid positions before attempting any solve. It's not elegant but it's what's available without reverse-engineering the NYT frontend every week. The tool I ended up using for this isn't publicly hosted. You'd need to set it up locally with Node.js and Puppeteer. The dependency list is manageable but the Chrome version mismatch issue comes up often. If your Puppeteer version doesn't match your installed Chrome, the headless instance will crash during the XHR interception phase. Check your versions with chrome --version and match the puppeteer-core package accordingly. Here's what this approach doesn't handle well. Themed Sunday puzzles with unusual wordplay clues will almost never solve cleanly through pure pattern matching. The solver treats every clue as a straight definition, so puns, abbreviations, and cryptic-style constructions fail. You'd need a natural language understanding component for those, and that's a whole different project. For the weekday puzzles though, the method works consistently enough that I've cut my personal solving time from about 12 minutes down to under 90 seconds for a full auto-solve and review pass.
Get the Full Details
If you just want to solve puzzles faster without building anything, there are browser extensions that do a fraction of this, but they rely on hardcoded answer databases and become useless the moment a new puzzle drops. The scraper-based approach stays current because it pulls the actual clue text in real time. The tradeoff is that it requires more maintenance and you're operating in a gray area regarding the NYT terms of service. They don't actively enforce against personal use scripts but they could change their detection methods at any point.
