Codenames Is Secretly a Lesson in Cryptographic Thinking

Codenames has sold over 10 million copies since 2015. It won the Spiel des Jahres in 2016, the most respected award in board gaming. The designer, Vlaada Chvatil, is known for heavy strategy games like Through the Ages and Mage Knight. Codenames is his lightest game, and it is also the one that accidentally teaches cryptographic thinking.

The premise is simple. A 5x5 grid of words sits on the table. Two spymasters know which words belong to which team. They give one-word clues plus a number, trying to get their teammates to guess the right words without touching the opponent's words or the assassin. That is the entire game. It plays in 15 minutes and costs $20.

What makes it interesting from a cryptography perspective is the encoding problem. The spymaster has to compress a set of target words into a single clue word, then transmit a count of how many words match. The receiver has to decompress that clue back into a set of guesses. That is a channel coding problem. Claude Shannon described the math behind it in 1948, and Codenames makes you feel it across a table.

If you want to practice the associative thinking the game demands, our Codenames Analyzer clusters words by semantic similarity and suggests clues. It is a useful tool for spymasters who want to test their instincts against a machine.

Codenames was designed by Vlaada Chvatil and published by Czech Games Edition in 2015. The game supports 2 to 8 players divided into two teams, with one spymaster per team. The board is a 5x5 grid of 25 random word cards. Each spymaster sees a key card that assigns every word on the grid to one of four categories: red team, blue team, neutral, or assassin.

On a turn, the spymaster gives a clue consisting of one word and one number. The word is the associative link. The number tells the field operatives how many words on the grid relate to that clue. For example, the spymaster might say "Ocean: 2" to point at "Whale" and "Wave." The operatives discuss and touch cards. A correct guess lets them keep guessing up to the stated number plus one. A wrong guess that hits an opponent's word ends the turn and scores for the other team. A wrong guess that hits the assassin ends the game immediately with a loss.

The first team to identify all their words wins. The assassin mechanic is what creates the tension. If your clue accidentally connects to the assassin word, a single misguess by your teammate kills you. This is why spymasters spend minutes staring at the grid before speaking. The cost of a bad clue is not a lost turn. It is the game.

The game won the Spiel des Jahres in 2016, the German game of the year award that has been running since 1979. Winning the Spiel des Jahres typically drives sales into the millions, and Codenames was no exception. It became a fixture in board game cafes and office game nights within a year of release.

The spymaster's job maps onto a classic cryptographic problem. You have a set of plaintext targets. You need to encode them into a compact message that your intended receiver can decode, while ensuring an adversary (the other team's operatives, who hear the same clue) cannot decode it to their advantage. The adversary is listening to every transmission.

The constraint is brutal. You get one word and one number. That is your entire bandwidth per turn. You cannot add qualifiers, gestures, or tone hints. The rules explicitly forbid linking the clue word to any word on the board by position, spelling, or root. You cannot say "Beach: 3" if "Beachball" is on the grid, because the clue word shares a root with a card. The rules also forbid using a number as the clue word, and the clue word cannot be a word on the grid.

This is a lossy compression problem with an adversarial listener. The spymaster is choosing a codeword that maximizes the probability that the receiver maps it to the intended targets, while minimizing the probability that it overlaps with the opponent's words or the assassin. If you have ever worked with error-correcting codes or channel capacity limits, the structure is familiar. You are fighting noise, and in Codenames the noise is the semantic ambiguity of language itself.

The Codenames Analyzer on this site does this computation explicitly. It uses word embedding models to measure semantic distance between the clue and every word on the grid, then ranks guesses by similarity. Playing against the analyzer gives you a sense of how a machine solves the same problem the spymaster faces, and where human intuition still beats it.

The core skill in Codenames is semantic clustering. You look at your target words and try to find a concept that binds several of them together without binding any opponent words. If your targets are "Saturn," "Mars," and "Comet," the clue "Planet: 3" seems obvious. But if "Earth" is an opponent's word, "Planet" is dangerous because your teammate might guess Earth first. You need a clue that captures the celestial bodies but excludes the one that belongs to the other team.

This is where ambiguity becomes the enemy. Every word in English carries multiple senses. "Bank" means a river bank and a financial institution. "Match" means a game, a fire starter, or a pairing. The spymaster has to pick clue words where the intended sense is the one the receiver will grab first. If the receiver latches onto the wrong sense, they guess the wrong word.

Good spymasters develop a feel for how their specific teammates associate words. This is not general semantic similarity. It is personal semantic similarity. If your teammate is a marine biologist, "Current" will read as ocean current before it reads as electrical current. If they are an electrician, the reverse is true. The best Codenames teams are not the ones with the best vocabulary. They are the ones with the best shared context.

This is the same problem that makes natural language a terrible cryptographic channel. Shannon pointed out that natural language is redundant and ambiguous, which is great for human communication over noisy channels but terrible for secure encoding. In Codenames, you are forced to use the worst possible channel for a cryptographic task, and that is what makes the game hard.

You can see the same clustering logic at work in our Frequency Analysis tool, which groups letters by how often they appear in a ciphertext. The principle is the same: find the pattern that binds things together, then use it to narrow down the answer.

Claude Shannon published "A Mathematical Theory of Communication" in 1948 in the Bell System Technical Journal. That paper founded information theory. It introduced the concept of channel capacity, which is the maximum rate at which information can be reliably transmitted over a communication channel with a given noise level. The Wikipedia article on Claude Shannon covers the broader impact, but the 1948 paper is the relevant source here.

Codenames is a live demonstration of channel capacity. The spymaster's channel is one word plus one number per turn. The noise is the semantic ambiguity of the English language and the presence of opponent words that could intercept the signal. The spymaster is trying to transmit as much information as possible (cover multiple target words in one clue) without exceeding the channel's capacity (the receiver's ability to correctly decode the clue).

When a spymaster gives "Ocean: 3," they are betting that the receiver will decode three words from one clue. That is a high-bandwidth transmission. If the grid has "Whale," "Wave," and "Shark" as targets, the bet is reasonable. If the grid also has "Beach" as an opponent word, the noise increases and the channel capacity drops. The spymaster should reduce the count to 2 or pick a different clue.

Shannon's theorem says that reliable communication is possible as long as the information rate stays below channel capacity. In Codenames, this means there is a maximum number of words you can safely claim per clue given the semantic noise on the board. Push past that and your teammate guesses wrong. The game forces you to estimate channel capacity on the fly, using nothing but your knowledge of words and your read on your teammate's brain.

This is why Codenames feels harder than it looks. You are doing real-time information-theoretic optimization with a noisy channel and an adversarial interceptor. The fact that it is packaged as a party game does not change the underlying problem.

People underestimate Codenames because the rules take two minutes to explain. The difficulty is not in the rules. It is in the encoding.

Consider a grid where your targets are "Piano," "Pitch," and "Note," and the assassin is "Score." The clue "Music: 3" seems safe. But "Score" is a musical term. Your teammate might grab "Score" before "Pitch," and the game ends. You need a clue that hits "Piano" and "Note" without touching "Score." "Keys: 2" works for Piano, but does it hit Note? Maybe, if your teammate thinks of musical keys. Or they might think of door keys and guess something unrelated.

Every grid produces these knots. The spymaster is constantly trading off coverage (how many words the clue reaches) against safety (how many wrong words the clue risks). A conservative spymaster gives "1" clues every turn and wins slowly. An aggressive spymaster gives "3" clues and either wins fast or loses to the assassin. The optimal strategy depends on the board state, the score, and how well you know your teammate's word associations.

The other thing that makes the game hard is that you cannot correct your teammate in real time. Once you give the clue, you sit in silence. If they reach for the wrong card, you watch. You cannot say "no, not that one." This is the cryptographic equivalent of a one-way channel with no feedback. Shannon's model assumes you can add error correction, but in Codenames the only error correction is the clue itself. You have to get it right before transmission.

If you enjoy this kind of pressure, the Mastermind Solver covers a different deduction game with a similar one-way information problem. And the Fallout Terminal Hacking tool tackles a word-guessing puzzle that shares Codenames' core loop of eliminating wrong answers from a fixed grid.

Codenames has spawned several variants. Codenames: Pictures (2016) replaces words with images, which changes the clustering problem entirely. Images have fewer obvious semantic links than words, so the spymaster's job shifts from language association to visual association. Codenames: Deep Undercover (2016) is the adults-only version with profanity and innuendo. Codenames: Duet (2017) is a two-player cooperative variant where both players are spymasters, each seeing only half the key. Duet is arguably the hardest version because both players are encoding and decoding simultaneously with partial information.

Codenames: Marvel (2019) and Codenames: Disney (2020) reskin the game with licensed properties. The word lists change, but the mechanics are identical. The Marvel version is interesting because the word pool is narrower (character names, locations, objects from the MCU), which reduces ambiguity and makes high-count clues more viable.

There is no formal competitive Codenames scene on the scale of Magic: The Gathering or Chess. The game is too dependent on teammate chemistry and shared vocabulary for standardized tournament play. But it is a staple at board game conventions, and unofficial tournaments run at events like Gen Con and PAX. The format is usually two-person teams with a time limit on clues, which pushes the game toward speed rather than depth.

The game's lasting influence is that it made word association a spectator sport. Before Codenames, party games were either trivia (you know it or you do not) or social deduction (you are lying or you are not). Codenames introduced the idea that the interesting moment is watching someone think about words, and that the tension comes from the gap between what the spymaster meant and what the receiver hears. That gap is the same gap that every cryptographic system tries to close.