How I Came to Build Pairdle and Accidentally Made It EPIC

I played Wordle every morning for a long time before I ever thought about building anything.

It deserves the praise. One puzzle a day, six guesses, five letters, and the same word for everybody on the planet. There is no energy meter, no shop, no tutorial. You cannot grind it. When you are done you are done, and you wait until tomorrow like everyone else. That restraint is the whole design, and it is much harder to pull off than it looks.

The trouble, for me, was that I wanted more of it. Not more puzzles a day — more thinking per puzzle. Once you learn a decent opener and get a feel for letter frequency, a lot of Wordle days resolve into bookkeeping. I wanted the part where you sit and stare to last longer.

Everything became an -ordle

I was obviously not alone, because the whole internet had the same reaction at once and the suffix escaped into the wild. Quordle put four boards on screen at the same time. Octordle went to eight, Sedecordle to sixteen. Nerdle swapped letters for arithmetic, Worldle for country shapes.

I want to be careful here, because two things are true at once and people usually only say one of them. Several of those games are genuinely good. Quordle in particular is well made and I have played a lot of it, and we happily recommend a few of them in our roundup of the best daily word games. They are also, to my mind, blatant copies. Take the five-letter grid, the six guesses, the green and yellow tiles, the one-a-day cadence and the shareable spoiler-free result grid, then run the identical loop four times instead of once. Being enjoyable and being derivative are not opposites. Most of that wave changed the quantity of Wordle, or the alphabet Wordle was played in, and left the actual mechanic exactly where they found it.

The bottom of the wave dropped the pretence entirely. Early in 2022 the App Store filled with apps that were simply Wordle with the serial numbers filed off, some charging money for it, at least one charging a subscription. Apple ended up purging them. It was a strange thing to watch, because the game being copied was free and took two minutes to play.

I did not want to add to that pile, at either end of it. Partly on principle and partly on ego: taking someone else's mechanic and changing the colours is not a fun way to spend six months, and I would rather fail at something of my own than succeed at somebody else's.

So I set myself a rule that turned out to be a good one. Whatever I built had to be different at the level of the mechanic, not the theme and not the quantity. Not Wordle with animals. Not Wordle four times over. A different thing to actually do with your hands.

Pairs were already on my mind

I had a head start I did not appreciate at the time.

While I was finishing a biology degree at the Harvard Extension School, I took CS50, and my final project was a word game called Pairagrams. It was not commercial and it was not polished, but I had spent weeks with my head inside the specific problem of letter pairs — how they behave, which ones are common, how differently a word reads when you stop seeing it as a run of single letters.

So when I started sketching, I did not begin from "what could I change about Wordle." I began from the thing I had already been chewing on. What if you never typed a letter at all? What if the atom of the game was two letters stuck together?

That is the entire idea, and the moment I drew it I knew it was a real game rather than a re-skin. A six-letter word is not six decisions. It is three. But they are much harder decisions, because you are no longer asking "is there an E in this word" — you are asking "does this word contain ND, and if so, does it sit in the middle or at the end?" Your reading brain does not natively work that way, which is exactly why it is fun. I have taken to calling it Word Sudoku, for reasons I will come back to at the end.

Why six letters, and then eight

I tried four letters first. Four letters is two pairs, and two pairs is not a puzzle — it is a coin toss you win on the second guess. There is nowhere to hide.

Six was the sweet spot. Three pairs gives you a genuine middle position, which is where most of the difficulty lives, and six-letter words are common enough that the answer never feels obscure.

Eight came later, for people who found six too comfortable, and it is a bigger jump than it sounds. Eight letters is four pairs, and a fourth position multiplies the arrangements you have to rule out. That became Pairdle Pro, which runs the same rules one size up.

Pro is the one I play. With six letters I feel I ought to be able to find the word — to read it straight off the pool rather than work it out by position. With four positions that stops being possible. Either you see the word immediately or you do not, and even when you think you see it, it is often not the word.

Attempt one: the board that gave itself away

Here is where the actual engineering started, and where I was wrong twice before I was right.

The answer is three pairs. But three tiles is not a game — you need a pool to hunt through. My first instinct was the obvious one: take the answer's three pairs, then fill the rest of the board with randomly generated two-letter combinations.

There are 676 possible two-letter combinations. Pulling from all of them produces a board like this.

ZP. QX. JV. Nobody has to think about those for even a second. You do not eliminate them, you just never look at them, so half the board is decoration. Worse, the tiles that do look like English quietly announce themselves as the answer. The puzzle solves itself by process of "which of these could possibly be in a word."

A pool is only difficult if every option in it is plausible.

Attempt two: eight real words

The fix was to stop generating pairs and start harvesting them. Take eight real six-letter words, slice each into its three pairs, and put all of it on the board.

Twenty-four tiles, every one of them lifted out of a real English word, so every one of them is a credible candidate. This was the version that first felt like the game I had in my head, and early Pairdle worked exactly this way.

Then I hit the bug that took the longest to think my way out of.

The bug: two tiles wearing the same face

Slice enough real words and pairs start colliding. FLOWER ends in ER. So does SILVER.

Now the board has two ER tiles that are visually and functionally identical, and the game has an ambiguity problem. If you place an ER and it comes back wrong, which ER did you eliminate? The answer is neither, because they are the same two letters — but the game is tracking them as separate objects, and the moment two objects are interchangeable, the bookkeeping underneath falls apart. Feedback stops being clean. Elimination stops being reliable. The one thing a deduction game absolutely must do is tell you the truth about what you have ruled out.

My first fix was to just delete the duplicate. That works, but it leaves the board a tile short, and do it a few times and the pool shrinks unevenly from day to day.

The fix, which is also the accident

The version I shipped does something slightly different, and it is the part of Pairdle I am fondest of. When a pair turns up more often than the answer actually needs, that copy is not deleted. It is swapped for a different pair drawn from a list of the most common letter combinations in English — the TH, ER, IN, ST, CH end of the alphabet, the pairs your eye accepts without question.

The ambiguity is gone, the board stays a constant size, and the replacement is just as plausible as everything around it.

And then the thing I did not plan happened.

Because those substitutes come from outside the eight source words, the finished board ends up holding letter pairs that were in none of the words I fed it. I built that step to fix a bookkeeping bug, and as far as I could tell it had also made the boards a little richer than my own recipe. I was quietly pleased with it. I called it a happy accident and I have told people about it more than once.

Hold that thought.

The exception I nearly shipped over

There is one more wrinkle, and it is the kind of thing that only shows up when a real player finds it.

"Remove the duplicates" is the wrong rule. Some answers need their duplicates.

BANANA is BA + NA + NA. If the generator strips duplicate pairs on principle, the board is missing a tile the answer requires, and the puzzle becomes unsolvable — the worst possible bug in a daily game, because everybody hits it on the same day. Words like EDITED, CHURCH, ICONIC and SEEDED all have the same shape.

So the rule had to get smarter. Count what the solution actually requires first, keep exactly that many, and only substitute the surplus. It is one extra line of thinking and it is the difference between a puzzle and a support inbox.

Why I call it Word Sudoku

The closest comparison I have found for Pairdle is not another word game at all.

When I play Sudoku I start pencilling numbers in, follow them until I hit a dead end, then back out and try a different pattern. Pairdle is that same motion. You put ER in the LAST position because ER is a common ending, and then you find you cannot build anything around it — or you press guess and it comes back grey, and off you go looking for another pattern. The pencilling-in is the game. It uses a different part of your brain than hunting for single letters does.

That is also where the two games genuinely part company, and it is the thing I am proudest of.

In Wordle you cannot really solve the first guess. There is nothing on the screen to reason from yet, so an opening solve is luck or it is cheating. In Pairdle the entire pool is in front of you from the first second, which means a one-guess solve is something you can actually deduce your way into. Most days you will not, and the hard ones are genuinely hard — on the two most punishing boards we have ever shipped, BADASS and REFUTE, players who solved them needed 4.4 guesses on average, and some boards will take you all six. (Those figures come from spring 2026 — read the next section before you take them as current.) But the possibility is real rather than a lottery ticket, and that changes how the first minute feels. You are reading, not stabbing.

Pro raises that ceiling again. One guess at four positions is genuinely hard and I miss far more than I hit. For real puzzle aficionados, I think Pairdle Pro is the true morning word challenge. Wordle gives you six guesses at a five-letter word, which is a generous allowance — if you are still missing at six, it usually is not the puzzle's fault.

NEWS FLASH — 8 August 2026

I was most of the way through writing this article when I decided to check a number.

I wanted to say how many words a typical board can spell, so I replayed the generator over every puzzle we have ever built. The answer came back: about five. That felt low. It is a board made from eight real words — five felt much too low.

So I looked at what the generator was actually producing, tile by tile, and found this.

It was substituting twenty-one of the twenty-four tiles. Every puzzle. Since December.

Not the duplicates. Almost everything. Of the eight words I so carefully harvested each day, the board was keeping the three pairs that spell the answer and throwing the other seven words in the bin, replacing them with generic filler off the common-bigram list. The decoys a player spends six guesses reasoning about were not decoys at all. They were wallpaper.

The cause is one word long. The code counted how many times each pair is required, then substituted any copy beyond that allowance — which is exactly right, except the allowance was built from the solution word alone and defaulted to zero. So every pair that was not in today's answer had an allowance of nought, and the very first time the generator met it, out it went.

const required = requiredCounts.get(letters) || 0;   // any non-solution pair -> 0
if (seen > required) { substitute(); }               // ...so 1 > 0, swap it

The fix is Math.max(1, ...). Keep the first sighting of every pair; let the answer's own needs raise the allowance where a word like BANANA calls for two. One word. It shipped on 30 December 2025 under a commit message that reads "fix paridle missing letter pairs", with a comment above it correctly describing the behaviour it did not have.

Nothing failed. That is the part I keep turning over. No test broke, no error surfaced, and the board still looked completely plausible — because filler drawn from the most common letter pairs in English is, by construction, exactly as convincing as the real thing. It took seven months and a spare afternoon fact-checking my own blog post.

Here is what the fix changed, measured across all 1,124 standard puzzles:

	before	after
tiles replaced with filler	21 of 24	0.84 of 24
boards needing no substitution	none	45%
words a board can spell	median 5	median 22

So the honest version of this article's title is that I did not accidentally make it epic. I accidentally made it thin, told myself a nice story about why, and only caught it because a number looked wrong and I could not leave it alone.

The good news is that the game you play today is the one I designed in the first place. The board is genuinely made of eight real words again, the decoys are real decoys, and it turns out those pieces recombine into far more words than I ever put in — around twenty-two on a typical board rather than five. That part was never the substitution's doing. It was the eight words all along.

If you have been playing since the spring: it got harder this week, and that is not your imagination.

What I would tell myself at the start

The mechanic came from a student project I had half forgotten. The hardest part was never the idea — it was the attempts I had to throw away, each of which looked completely reasonable until I actually played it.

And the thing I was proudest of turned out to be a defect I had been describing as a feature, which is a lesson I would rather have learned on somebody else's game. If there is anything to take from it: the bug that survives is never the one that crashes. It is the one that produces output you can tell a nice story about. I had the measurement in front of me months ago. What I did not have was the willingness to be suspicious of a number that flattered me.

If you want to see the finished version of all this, today's Pairdle is here, and there is a fuller walkthrough of the board in how to play.

I still play Wordle most mornings.