Make a clue game

Pick the single word that satisfies all four clues. They come at the term from different angles — synonym, function, category, contrast — which the data carries for over a million words.

Below are six real puzzles from the build this page counts. Tap an answer: every option is scored against every clue. The answer connects to all four; each wrong option has zero connection to at least one clue, and that zero is the player’s elimination path.

TAP AN ANSWER · ITS STRENGTH MATRIX

Strength: 100 − the clue’s position in the option’s ranked association cloud; not listed = 0. Separation: the answer’s total minus the best wrong option’s.

Problems

The answer has to match all four clues — which words carry four separate angles?
Wrong options have to be almost right — matching some clues, clearly missing others; where do they come from?
Everything on screen has to be familiar — five words show per puzzle; what pool do they draw from?
It has to stay family-safe — which filter, applied to which words?
The game has to last — how many puzzles does English hold?

Relevant data sources from Linguabase

This example draws on 6 of the 13 Linguabase tables — two carry the meaning structure, one splits it by sense, and three keep everything familiar, unrelated, and safe.

associations	A ranked cloud of related words for every word, strongest first. Clues come from the answer’s cloud; an option’s strength on a clue is 100 minus the clue’s position in its cloud.
senses	The canonical senses of each word. A puzzle needs an answer with at least two.
sense_associations	A separate cloud per sense. Says which sense each clue belongs to, so the four clues can be forced to spread.
vocabulary	A familiarity rank for every word. Answers come from the top 10,000; wrong options from the top 30,000.
word families	Which words are relatives. Keeps BARKING out of BARK’s clues and out of its wrong options.
scores & labels	A content rating and a sensitivity score per word — the family-safe filter for everything shown.

Every count on this page is measured against the June 2026 public release; this page draws answers from the 10,000 most familiar words.

Example 1: Can the game yield a lot of puzzles?

Below, the 10,000 most familiar words are cut down to the ones that can anchor a puzzle — four clues across senses, three clean wrong options, a 100-point win. Six steps, run one at a time; scrub the survivors at every stage. Step 7 re-runs everything at other pool sizes.

PROCESSING QUEUEDATA INSPECTOR

NEXT STEP

DATA INSPECTOR

Scroll the output:

Multi-word entries ride through every step as ordinary words: ICE CREAM anchors a puzzle whose wrong options are SNOW CONE, FROZEN YOGURT, and CHOCOLATE MILKSHAKE.

Example 2: Build one puzzle

Five steps take one word to a shipped puzzle: fetch its cloud and senses, pick four clues that spread across senses, then hunt wrong options among the words that share two or more of those clues. Push TOLL through and watch the puzzle record build on the right; then swap in SEAL — whose clues run to PINNIPED and GASKET — and GRANDMOTHER, which never clears the first step.

BUILD QUEUEPUZZLE RECORD

NEXT STEP

PUZZLE RECORD

SEAL fails: its senses are tidy — animal, authentication, closure — but its strongest associates are specialist words, and too few familiar words share two of them. The build wants an answer whose separate senses all reach into common vocabulary.

Example 3: What a puzzle carries at play time

Nothing is computed at play time: a puzzle rides as one small record — the answer, the four clues, the three wrong options, and the full strength matrix. The matrix is what feedback reads: after a wrong pick, the game can name the zero. HIGHWAY, picked in the TOLL puzzle, has no connection to DEATH or CHIME.

THE PUZZLE PACK · ONE RECORD

All puzzles together are KB minified; a month of dailies is KB.

Outputs

At this point the supply is counted (Example 1), the build is walked (2), and the pack is written (3). What each table contributed:

	Example 1 (counting)	Example 2 (the build)	Example 3 (the pack)
associations	clue supply, decoy matching	the cloud, every strength	the strength matrix
senses	the 2+ senses step	the sense count	—
sense_associations	the clue-spread step	which sense each clue is from	—
vocabulary	the 10k pool, the 30k pool	candidate ranks	—
word families	relative screens in both picks	the unrelated check	—
scores & labels	the family-safe step	—	—

Other considerations

5,488 anchors is one spec against one release. Still open:

Anchors undercount puzzles — this page picks ONE clue set per answer, deterministically. An answer with six senses supports several distinct puzzles; 5,488 is a floor, not the supply.
The mechanics don’t judge answer quality — NOT and YES both anchor valid puzzles at this spec: four clues, clean wrong options, a 100-point win. Curation stays a job; scores & labels carries game_suitability and concreteness for sharper cuts than this page makes.
Strength is a position proxy — 100 minus a list position, not a semantic measurement. Near the horizon (position 99) a “match” is faint; the 100-point separation floor is what keeps faint matches from deciding puzzles.
Senses come at two grains — this build reads only the canonical labels in senses; sense_associations also carries hundreds of finer senses per word (SEAL alone has 412 rows). A finer spread rule is available if a puzzle feels samey.
Sensitivity is high = safe — the filter here keeps content_rating G or PG and sensitivity above 3; 5 is the neutral majority value. Production games read more of scores & labels than this one filter.
File mechanics — the downloads are not rank-sorted and the TSVs carry a header row. Association lists are parquet list<string> columns; in the TSVs they are pipe-delimited — split on |.

Next steps

For clue games: download the tables from any of the three repositories, hypothesize your own clue design, and count it the way Example 1 counted this one — then build out the scripts in your language of choice. The walkthrough demos and each table’s README on the mirrors go deeper — and this page itself is a workable hand-off to whoever builds for you.

To help, two heavily documented draft artifacts for you or your coding agent to continue from:

linguabase-clue-game.py — the full build, ready to run against the tables; writes puzzles.tsv and puzzles.json. Opens a line-numbered view of the script, with the download.
clue-game-llm.txt — tables, method, measured counts, and the script in one plain-text file an LLM can work from.