wordlb0t

When you play Wordle, the computer picks the word and you guess. Wordleb0t flips the script and guesses the word that you choose.

Play it here, or follow @wordleb0t on X for a daily post of its guesses against that day's NYT puzzle.

Introduction

Start by picking a secret word (between 4 and 11 letters).

Wordleb0t gets 6 tries to guess your word. Each time, you can check the guess by clicking on the letters to change their background color. Green means a letter is correct in this spot; yellow means a letter is elsewhere in the target word; gray means a letter is not in the target word at all. Once you've colored in every letter, click the checkmark to lock-in your feedback and Wordleb0t will make its next guess.

If you realize you made a mistake, there's an undo button to go back. Speed things up with the number keys on your keyboard- each one will increment the color of the of the associated letter.

Source

The core of this repository was forked from @chordbug's excellent hello-wordl project, including @lynn's initial editing of Peter Norvig's word frequency list (derived from the Google Books Ngrams dataset).

Wordleb0t was released in January 2022, several months before the New York Times launched its own Wordle-solving analysis tool in April 2022.

Algorithm

At every turn, Wordleb0t narrows the field to the words consistent with every clue you've given so far - the right letters in the right (or wrong) spots, and the letters ruled out entirely.

When that field is small (3 or fewer words), Wordleb0t just guesses the most promising word left in it, ranked by how common its letters are and, for words that have actually been NYT Wordle answers before, how likely NYT is to reuse it.

When the field is wider than that, ranking candidates against each other stops being enough: a cluster of words that share almost every letter (e.g. wafer/wager/hater/later/eager/...) can trap a guesser into testing one look-alike per turn, since no candidate word can distinguish more than one of the unconfirmed letters at a time. Wordleb0t instead picks a guess - which may not even be a possible answer itself - that best splits up the remaining field, using Shannon entropy to measure how much a guess narrows things down: for each candidate guess, Wordleb0t simulates the clue it would get back against every remaining word, groups the results by that clue pattern, and scores the guess by how evenly and finely that groups the field. A guess that's equally likely to produce any of several very different clues carries more information (higher entropy) than one that mostly produces the same clue no matter which remaining word turns out to be the answer, so it's expected to eliminate more of the field in one go regardless of what comes back. Ties go to an actual candidate word, since guessing one also has a chance of winning outright.

The best guess for a cluster of look-alikes is often a globally uncommon word - splitting up bound/found/hound/mound/pound/wound calls for a word like lymph, not a more familiar one - so every dictionary word of the right length is a scouting candidate, not just common ones. Scoring every candidate against every remaining word is the expensive part, though, so how many candidates get considered shrinks as the remaining field grows: a handful of remaining words can afford to search the whole dictionary, while hundreds need a smaller search to stay responsive.

None of this is worth it on the last guess you're allowed, though: a scout that isn't a candidate has no chance of being the answer, while any remaining candidate does, so with one guess left Wordleb0t always falls back to guessing its best remaining candidate outright.

This is the same information-theoretic idea behind 3Blue1Brown's "Solving Wordle using information theory" (writeup here), adapted to run efficiently for words of any length and to weigh a candidate's own chance of being the answer against pure information gain. The discussion that led to it is in issue #31.

For developers

If you can make Wordleb0t smarter, please feel free to fork the code on GitHub or leave a pull request.

To run the code locally, first install Node.js and npm. Then, in this directory, open a terminal and run npm install followed by npm run start (npm run dev is an alias). hello wordl will be running at http://localhost:3000/. Any changes you make to the source code will be reflected there. Have fun!

Finally, npm run deploy will deploy your code to the gh-pages branch of your fork, so that everyone can play your version at https://yourname.github.io/wordleb0t (or the name of your fork if you renamed it).

Follow-up worksheets

If you always open with the same first guess (say, adieu or slate), you can generate an SVG worksheet covering every possible gray/yellow/green result for that guess, paired with Wordleb0t's suggested next guess for each one:

npm run worksheet -- adieu

This writes worksheet-adieu.svg to the current directory (pass a second argument to choose a different output path). Convert it to PNG or PDF with any SVG tool, e.g. Inkscape or rsvg-convert.

The all-gray and all-yellow results are shown as two cards next to the title, since each is just a single result rather than a whole column's worth. The remaining results are grouped into columns by how many letters came back colored (yellow or green) and by which of those are green vs. yellow, so you can jump straight to the column that matches what you saw. Results with 4 or 5 colored letters are broken out into their own section below the rest, since there are far more color arrangements than there are worthwhile columns to spread them across.

Daily X post

The Post Daily Result GitHub Actions workflow runs on a schedule, fetching that day's NYT Wordle answer, solving it with Wordleb0t, and posting the result to @wordleb0t. It reads Twitter/X API credentials from repo secrets (TWITTER_API_KEY, TWITTER_API_SECRET, TWITTER_ACCESS_TOKEN, TWITTER_ACCESS_TOKEN_SECRET) and can also be triggered manually from the Actions tab.