Codenames, but the Bot Only Knows the Board

About eight years ago, a Codenames game came down to a single turn. My team was almost certainly going to lose unless we correctly guessed all of our remaining words. The clue wasn't enough to get us there, but a teammate pointed out that, based on the layout of the words on the 5x5 grid, one of our candidate picks would have produced a pattern that was "too weird." We followed that reasoning, guessed the last word based on position, and won.

That experience stuck with me. What if someone simply memorized all 40 setup cards that ship with the game? How strong could a player be if they ignored word meaning entirely and played only from the spatial patterns those cards produce?

If you're not familiar with Codenames: two teams compete on a 5x5 word grid. Each word is secretly assigned to one of the blue team, the red team, a neutral tile, or the assassin. The spymaster on each side knows the full assignment and gives single-word clues to guide their teammates to their own words while avoiding the assassin and the opposing words. The first team to guess all of its words wins. Playing with the physical cards is not strictly necessary — any board is a valid configuration as long as the word counts and turn order are respected — but the 40-card deck is the standard setup. A specific pattern geometry underlies it: the start team always gets exactly 9 words, the other team 8, plus 7 neutral and 1 assassin. The board rotates through 4 orientations, yielding 160 unique layouts from the 40 cards.

The Bot’s Only Weapon: Pattern Memory

A teammate hadn't memorized the cards, but after enough plays had built an intuition for which configurations the deck allows. The earlier winning guess was later confirmed when we checked the board cards afterward. I've noticed players guessing an extra word on the first turn just to mess around, which works more often than not — since they have no information, that success rate suggests some unconscious pattern use is happening, whether from the board geometry or by noticing what the spymaster is looking at.

When you look at the actual set of board layouts, a few things stand out: no card has a full 5-in-a-row or column. Only the start player's color (blue) ever reaches 4 in a row, and if that row sits on an edge, the 5th position in that line is forced to be a red word. Six connected blue tiles never appear, but there is a card with a red cluster of 6, and another with two pairs and a 4-run centered next to a 2-run. Even if no one consciously tracks these facts, repeated play makes certain geometries "feel right" and others feel obviously wrong. This gap between conscious knowledge and unconscious bias is exactly the side channel a pattern-only bot can exploit far more systematically than a human ever will.

The demo bot in this article plays under an artificial disadvantage: it only ever acts when its deduction space collapses to one possible board. It wins instantly at the start of its turn if all remaining matching boards agree on its words. It doesn't guess early to reduce the search space, and it doesn't believe you'll make an error; the most naive possible "memorized board set" bot wins on its turn if you've left only one plausible layout standing. Trying it against human players showed that in most games, without forgoing the bot's rules, the bot still won against players of typical skill. If a human guesses any three words a turn with perfect accuracy every single time, the bot nearly always deduces the full set of possible boards down to one after a single turn. The only meaningful handicap versus a human is that the human starts with no information about which layouts are geometrically possible — unless they toggle the "show remaining boards" display, which lets a person use identical knowledge. At that point, the game shifts from word association to a pure distributed constraint-solving problem.

How Real Is This Side Channel?

After investigating this, I don't think it poses a serious problem for Codenames in practice. Word-guessing games are not usually meant to be mechanically solved; most break once players are either very good or play to win rather than have fun. Codenames has more headroom than most because effective play requires learning a combinatorial set of word-association clues, not a small fixed list. Games like Taboo or Just One run dry once a group has cycled through the deck once or twice; players start cluing around the forbidden words out of rote memory rather than fresh thought. With Codenames, the barrier to "solving" the word-association space is far higher. The positional issue from the 40-card deck is real but subtle relative to those problems, and it's avoidable entirely by using a digital randomizer for board generation — something no group I've played with has ever bothered to do. Even if people don't consciously notice the pattern constraints, repeated play does bias their judgment. After a couple dozen games, I'm fairly confident some subconscious geometric filtering was already affecting my calls. But the repeated-word problem in other games is far more direct; that's why the Codenames variant occupies a better spot for replayability.

Appendix: AI as a Naive Programmer

For this post I used Storytell and Cursor as a non-programmer would, prompting for code and running tests without inspecting the logic, then telling the assistant what broke until it worked. In that basic loop, the total time wasn't far off what hand-writing would've taken — maybe twice as long, but programmer estimates are unreliable, and possibly comparable. I feel less confident in the code's correctness than if I'd written it myself, yet the result that "works basically well enough" is still surprising — the user never needs to understand the code, only to recognize when output is wrong and ask for a fix. An experienced programmer fixing specific bugs interactively would get far more value.

I saw the same cycles over and over with both tools: an incorrect generation that its own execution could detect, a request to fix it, another generated attempt that doesn't typecheck or test, then a pass after a pasted error message. In one frustrating case, fixing one bug reintroduced a prior one even when explicitly asked not to. It's a shame that neither environment offers an execution sandbox that would catch simple invalids and attempt corrections automatically — they don't train or fine-tune with such a loop. When I asked Josh Bleecher Snyder why, he pointed to the broad industry embrace of "The Bitter Lesson": most labs believe that this kind of mundane environment work will be obsoleted by fundamental advancement. Sam Altman has advised companies building on OpenAI APIs to assume massive leaps and position accordingly, and conversations with founders and VCs suggest they've taken that to heart. As a layperson with no real ML work in over a decade, I'd still bet a mundane system built in the GPT-3 era that runs against code execution would be useful today for labeled data and improved model outputs. For the biggest labs, marginal labeled examples might not justify focus. As a portfolio move at that scale, though, it's a cheap hedge against the collective bet that raw scale will subsume this sort of tooling.

Does Playing to Win Ruin It?

If you ask Dave Sirlin, anyone who won't break a game within legal bounds is a sucker. I'll take the opposite view: play any way you enjoy, and find a group that agrees on that. In games like Perfect Words or Codenames-as-played-to-win, the exploitative route is about memorizing arbitrary board sets — no creativity or cleverness involved — which seems dull as a game activity. I also notice groups quietly pick up spatial intuitions regardless of intent; my strongest instinct is not that anyone cheats at Codenames, but that after enough cards, an adversarial geometric read becomes unavoidable for experienced players. The fix that preserves the full intended experience is digital board generation.

A Few Words on Other Word Games

This list only covers games predicated on word meaning — not mechanics like Banagrams or Scrabble, nor visual mapping like Dixit or Codenames: Pictures. It also deliberately skips what are essentially universal complaints, like a game being word-based or light in weight.

Taboo

Rapid-fire cluing with a static list of five forbidden words per target. Replayability bottoms out fast — a group can exhaust the deck in a couple of sittings, and players soon learn to describe around the block list.

Just One

Allies each write one clue, then all duplicates are removed. Because the forbidden list comes from the players each round, it's more replayable than Taboo by far, but a fixed word deck plus finite clever cluing approaches means my group was recognizing everything after a couple dozen plays. The physical components don't justify the $30 price next to Codenames' $20, and word difficulty is inconsistent — "grotto" is much harder to clue in English than in the original French.

Perfect Words

Brilliant core concept that the designers clearly didn't playtest — players who try at all can trivialize it within the first session. The game works only if everyone refuses to try the intended mechanical win.

Montage

A two-on-two clue-and-guess race with constraints on letter positions, and both the clue and target word must be guessed before opponents. It feels harder than anything on this list; first-time players commonly fail to give any solid clue. Strong decode capabilities give a large advantage, and new players generally won't win a single game without a deliberate handicap. That asymmetry is a turn-off for many, but it's exactly what I admire about it. If you want a no-excuses test of word skill, Montage is the recommendation.

Word Blur

Clues are built from a table covered in tiny word tiles. High replayability in principle, but out of print for a decade and genuinely hard to source at all. Playing Word Slam, its easier deck-based knockoff, is simpler mechanically and far less chaotic. Word Blur's search over a chaotic tile board is why some people find it alarming.

Codenames: Duet

The co-op variant works fine with more than two players if you accept quieter side discussion. Replayability sits at roughly Codenames level, since the 40 boards remain fixed.

Decrypto

Each player rotates giving a clue only their teammate should map to the right target, while opponents transcribe the clue and try to infer what it meant against the shared vocabulary. This one punishes the weakest link far more than any other co-op game — I've never seen a random-group game decided by anyone other than the low-skill player on one side or the other. Too little data on replayability ceiling from my end.

Person Do Thing

Taboo inverted: a list of allowed clues rather than forbidden ones. Extremely constrained verbal creativity, but the deck is embarrassingly small — a couple plays will exhaust it. My friend David Turner suggests it's like Concept in that a dedicated group will build a persistent convention language ("flag" meanings locked to particular colors, country modifiers attached) in very few sessions.

Semantle

Wordle for semantics over word2vec embeddings. Doesn't click for most; players who lack a mental model of the vector space see the scores as random noise. A small number of people will grind through hundreds of games to feel out the inference — but even those tend to burn out.