How to play Contexto

Contexto is a daily word game in which you try to find one hidden word, and the only feedback you get is a single number telling you how close each guess is in meaning. Not in spelling — meaning. There is no letter grid, no guess limit and no way to lose. You simply keep guessing until the number reads 1. If you want to try a round before reading any further, today’s puzzle is open now; this guide will make more sense with a few ranks already on your board.

The rules in sixty seconds

One word is hidden. You type any word you like and press enter. The game replies with that word’s rank: its position in a list of roughly 30,000 words sorted by how close each one sits in meaning to the answer.

• Rank 1 is the answer itself. Finding it ends the round.
• Rank 2 is the word that shares the most context with the answer across everything the model has read.
• Rank 24,000 is a word with essentially nothing in common with it.

That is the whole game. Guesses are unlimited, wrong guesses cost nothing but time, and everyone playing in your language gets the same word, which resets at midnight in your local time zone. Portuguese and Spanish run their own separate word lists, so the answer is genuinely different in each language rather than a translation of one original.

Two rules catch people out on their first few rounds. First, guesses are reduced to a base form before being ranked — running is scored as run, moons as moon — so you cannot farm extra information out of plurals or tenses. Second, spelling is never corrected. A typo is simply ranked as whatever nonsense word you typed, which usually lands somewhere past 20,000 and looks like a meaningful signal when it is not.

What the rank actually measures

The number comes from a word embedding, and understanding what one is turns Contexto from a guessing game into a solvable puzzle.

A language model reads an enormous amount of text and, for every word, records the company that word keeps: which words appear near it, in what sorts of sentences, how often. It compresses that record into a list of a few hundred numbers — a vector. Words used in similar ways end up with similar vectors. That is the entire trick, and it is why the technique is often summarised as you shall know a word by the company it keeps.

To build a day’s puzzle, Contexto takes the answer’s vector and measures the distance from it to the vector of every other word in its dictionary. It sorts that list, closest first. Your guess’s rank is just its line number in that sorted list.

Three consequences follow, and each one explains a whole class of “that can’t be right” moments:

The ranking reflects usage, not definitions. The model has never seen a dictionary. If two words appear in the same kinds of sentences, it considers them close, whatever a lexicographer would say. This is why anchor and harbour are neighbours, but the broadcasting sense of anchor — a news presenter — contributes almost nothing, so presenter sits thousands of places away.

Opposites are close, not far. Hot and cold, always and never, buy and sell: each pair appears in nearly identical sentences, so their vectors are nearly identical too. Guessing the opposite of your best word is a legitimate and often excellent move. It is also the single most common way players walk confidently in the wrong direction for twenty guesses.

Rank is not distance, and it is not linear. Because the ranking is a sorted list, the gap between rank 1 and rank 100 covers a tiny slice of meaning-space, while the gap between rank 10,000 and rank 10,100 covers almost none of interest. Improving from 8,000 to 4,000 sounds like halving the distance; in practice it means very little. Improving from 400 to 200 is enormous. Read big numbers as a category signal — wrong region — and small numbers as a gradient.

One more thing worth knowing: the dictionary is finite. Roughly 30,000 words are ranked, so proper nouns, very rare words and most compounds are simply not in it and will be rejected rather than ranked.

The colour bands

Contexto colours every guess, and the three bands are not decoration — they are the game’s actual difficulty structure.

The bands matter because they tell you which mode to play in. In red, you are exploring — the goal is coverage, and you should be jumping between unrelated areas of vocabulary as fast as you can think of them. In yellow, you are orienting — you have found a category and should be working out which part of it the answer lives in. In green, you are converging — every guess should be a small, deliberate step from your current best word.

The most common failure in the whole game is playing the wrong mode for your band: grinding out careful, incremental guesses while everything is still red, or throwing wild ideas around after you have already found a green.

A note on colour and accessibility: the rank number is always displayed alongside the colour, and it is the authoritative signal. Nothing in the game requires you to distinguish green from orange.

The tip system

When you are stuck, Contexto can give you a word. It does not tell you the answer; it tells you a word that sits somewhere useful in the ranking, and you carry on from there.

There are three tip difficulty settings, and they change what the game is willing to hand you:

• Easy — tips come from very close to the answer, often inside the top 20. One easy tip usually ends the round within a few guesses.
• Medium — tips land in the middle of the green band, roughly ranks 50–300. Enough to relocate you to the right neighbourhood without finishing the job.
• Hard — tips come from the yellow band. They tell you the general category and nothing more.

Which one to use is a taste question, but the practical advice is: change the setting before you get frustrated, not after. Players who leave it on Easy tend to reach for a tip at the first sign of difficulty and never build the skill; players who leave it on Hard tend to grind for an hour and stop enjoying it. Medium is the setting most regular players settle on.

Tips are also the cheapest way to learn the model’s quirks. Take a tip, look at its rank, and ask why the game considers that word so close. Doing that a few times a week teaches you more about how the ranking behaves than any amount of guessing does. If you would rather have prose clues than a word, the daily answer page publishes three progressive hints that narrow the field without naming anything.

Which words can be the answer

Knowing the shape of the dictionary saves guesses, because it rules out whole categories before you spend anything on them.

Roughly 30,000 words are ranked. That is a curated common-vocabulary list, not everything in English, and it excludes several things that players reach for constantly:

• Proper nouns. No countries, no brands, no names. Paris and Shakespeare are not in the list and will be rejected rather than ranked.
• Hyphenated and multi-word terms. Ice cream, self-esteem and check-in have no single vector, so they cannot be guessed as units.
• Abbreviations and initialisms. TV, NASA, DIY are out.
• Very rare or technical words. If you would not meet it in a newspaper, it is probably absent.

There is an important distinction between a rejected guess and a badly ranked one. A rejected guess — the game tells you the word is not in its dictionary — carries no information at all: it does not mean the answer is unrelated, only that your word is not in the list. A guess that comes back at 26,000 does carry information: that word is in the dictionary and is genuinely far away. Players routinely misread a run of rejections as “I am in the wrong area” when it actually means “I am using vocabulary the game does not have”.

The answer itself is drawn from a narrower pool than the ranked list: common, everyday words, weighted toward concrete nouns but with a regular supply of verbs, adjectives and abstract nouns. In practice, if a candidate feels slightly too ordinary to be a puzzle answer, that is a point in its favour rather than against it.

Inflected forms are folded before ranking, so the dictionary is smaller than it looks: run, runs, ran and running are one entry. This is why cycling through the forms of a word is four guesses for one piece of information, and it is worth internalising early.

What counts as a good score

There is no par, but there are useful reference points drawn from how days actually play out:

• Under 15 guesses — an excellent round. Usually means an early guess landed inside the top 50 by luck or by a very well-chosen opener.
• 15 to 40 — a good, normal round. Most experienced players live here.
• 40 to 80 — an ordinary round, or a hard word played well.
• Over 100 — either an abstract answer or a round where you committed to the wrong category early and did not test out of it.

The average across a typical week sits somewhere around 40. Individual days swing wildly: a concrete noun with a dense neighbourhood, like anchor or compass, can average under 30, while an abstract social noun like rumour can average over 60 with a much wider spread between the fastest and slowest players.

Guess count is worth tracking not as a score but as a diagnostic. If your counts are consistently high, the cause is almost always insufficient early exploration — you are converging before you have found the right region. If they are high only on abstract words, the fix is vocabulary rather than method, and the tips page has a set of opening words chosen to cover abstract space.

A full worked round

Here is a real round on the answer moon, played the way the method above suggests, with the actual ranks the model returns.

Exploring. Open with words from unrelated areas of vocabulary. The goal is not to be right; it is to find out which region is warm as cheaply as possible.

• coffee → 8,244. Red. Food and drink is cold.
• keyboard → 15,881. Red, and colder. Objects and technology are not it.
• ocean → 1,043. Yellow. Something here.

Three guesses in, we have a direction. Note that ocean being yellow does not mean the answer is nautical — it means the answer shares context with large natural things.

Orienting. Work the yellow signal. Test which part of that space is warmest by sampling widely within it rather than deeply.

• mountain → 6,742. Colder. So it is not “big natural feature” in general.
• night → 55. Green, and strongly. That is the real signal.
• star → 88. Green. Confirms it: we are in night and sky.

Six guesses in and we are inside the top 100. Notice that ocean was a red herring in category terms — it scored well because of tides, not because of water.

Converging. Now every guess is a deliberate small step, and the ranking is precise enough to trust.

• sky → 68. Slightly better than star.
• planet → 104. Slightly worse. So: not astronomy in general.
• satellite → 9. A big jump. The answer is a specific body, not a class of them.
• orbit → 12. Close, but satellite is still better.
• lunar → 2. One away.
• moon → 1. Done, in twelve guesses.

The shape of that round is the shape of most good rounds: three or four cheap wide guesses, a couple of guesses to identify the category, then a short precise walk inward. The strategy guide breaks the middle phase down into a repeatable procedure — it is the part where most guesses are wasted.

The six mistakes that cost the most guesses

Converging too early. Finding a yellow at guess two and spending the next thirty guesses inside that category. Before you commit, spend three guesses in completely unrelated areas. If one of them beats your yellow, you have saved yourself twenty guesses.

Reading red ranks as data. The difference between 9,000 and 14,000 is not information. Both mean wrong region. Players who chase small improvements in the red band can burn fifty guesses without moving.

Guessing variants of the same word. Run, running, runner, ran all reduce to nearly the same vector. Four guesses, one piece of information.

Ignoring opposites. If your best word is cold at rank 40, hot is very likely also near 40 — and the answer may be either one, or something between them. Testing the antonym is often the fastest single guess available.

Only guessing nouns. Verbs and adjectives are ranked too, and a well-chosen verb often locates an abstract answer far faster than a noun does. Soothe is what cracks a lullaby round; no amount of naming objects would.

Forgetting the tip system exists. A tip is not a failure. Used at the right moment — when you have a green but cannot improve on it — it costs nothing and teaches you something about the model.

Where to go next

If you have the mechanics and want the tactics, the tips page collects the tricks and opening words that shave the most guesses off an ordinary round. If you would rather have a full procedure to follow — how to sweep categories systematically, when to abandon a lead, when to take a hint — the strategy guide sets one out step by step.

For the specific questions this guide does not cover — who made Contexto, whether there is an app, why it is not loading — see the frequently asked questions. And when you want to check today’s board against the record, the answer page has the current puzzle’s hints and closest words with the solution kept behind a spoiler blur.

Otherwise: go and play today’s puzzle. The fastest way to internalise any of this is to lose one round badly and see exactly where the guesses went.