In a Nutshell is a group-the-words puzzle. In this genre, the player finds groups of related words, and every word fits exactly one group. Use the panel to reconfigure it on the fly — rows, words per set, vocabulary, word length. The Puzzle Data Sample tab shows the actual records behind whatever you’ve set up.
This example draws on 5 of the 13 Linguabase tables — one supplies the groups, two keep them apart, two pick and screen the words.
| topics | The group supply: 373,382 graded rows, each a short headword, a sense line, and member words — built for grouping play. |
| vocabulary | A familiarity rank for every word. Decides which members can sit on a tile, and orders a board’s groups from easiest to trickiest. |
| scores & labels | A content rating and a sensitivity score per word. The sixteen tiles all show on screen, so every word is checked, not just its group. |
| associations | A ranked association cloud per word — the overlap detector. A tile that sits in another group’s cloud gives that group away. |
| categories | 72,802 curated pools, category → members — the phantom detector. Three tiles from different groups that share a pool look like a fifth group. |
The supply starts as the whole topics table and ends as boards. Run the six filters one at a time and scrub the surviving rows at each stage; step 7 re-runs the count with the tile words capped shallower or deeper.
Four groups from 42,314 is 1.3 × 1017 combinations on paper. That number is not a supply: a daily game retires each group after one use, which caps the supply at 10,578 boards — 29 years — and the checks in Example 2 reject most pairings before they reach a board. The measured floor: reading the queue top-down, the first 4,000 groups seat 844 boards — a board a day for 2.3 years — with 38,314 groups still unread below them.
A board fails in play the moment one word fits two groups. Five checks run against every candidate group before it seats:
Run over every pair from those top 300 groups, the checks pass 16,260 of the 44,850 pairings; the phantom-pool check does most of the rejecting.
Below is the demo’s first board: three groups seated, one slot open. The queue offered 96 candidates for that slot; 95 failed. Tap a candidate to run its checks — the offending words are named.
The five shown rejections stand for the other 90, which fell the same ways — 61 phantom, 14 pool, 8 cloud, 7 tile.
The curation of word groups is essential for word categorization games, but you will need handcrafted puzzle-making to revise and polish your levels. The same groups, once they pass these tests and your handcrafting, drive a completely different mechanic.
Compare here the same kind of categorization puzzle data, manifested in a ball-blast puzzler that, like NYT Connections, asks you to find the hidden groups — then adds spatial reasoning, since the words tumble and collide.
Players drag through “jerk,” “tug,” “snatch,” and “wrench” once they notice these are all ways of pulling. The data supplies those groups, balanced so each one holds together and no word fits two at once.
The builder reads a queue — best quality first, then the groups whose four tiles run most familiar — and seats the first four groups that clear the checks. A rejected candidate goes back to the queue for a later board; a candidate that shares three of its four tiles with an already-seated group is skipped as a near-twin. Run the five steps and watch the board record fill; board 1 took 139 candidates. The full script follows under Next steps.
Tap four words that belong together, then check them:
Every count on this page is measured against the June 2026 release — the same files on the three public mirrors. The demo’s eight boards come out of the build.
A board ships as one small record: the sixteen tiles in their shuffled arrangement, and the four groups in playing order. The reveal a player sees on solving a group is the topics row itself — headword as the name, sense line as the explanation. Nothing is written by hand.
The four groups are ordered by the mean familiarity rank of their tiles: the group whose words run commonest plays first. On board 1 that puts I’m home (tiles averaging rank 1,142) first and piranha (1,960) last. The table also grades each group’s obviousness, 1 to 5; every group on the demo’s eight boards grades 5, so rank does the ordering here.
All 844 boards written as JSON weigh 624 KB; the eight embedded in this page average 746 bytes apiece. Play time is comparisons, nothing more: match the player’s four picks against the four groups, count the overlap for the “one away” message, read the reveal off the record.
At this point the supply is counted (Example 1), the checks are defined (2), the boards are built (3), and each board ships as a record (4). What each table contributed:
| Example 1 (counting) | Example 2 (exclusivity) | Example 3 (the build) | Example 4 (the bundle) | |
| topics | the supply itself | full member lists feed check 2 | the queue | name + sense are the reveal |
| vocabulary | the tile window, rank 1,000–30,000 | — | orders the queue | orders the groups |
| scores & labels | screens every tile candidate | — | — | — |
| associations | — | checks 3 and 4 | rejects candidates mid-scan | — |
| categories | — | check 5 | rejects candidates mid-scan | — |
844 boards is one spec against one release. Still open:
For categories games: download the tables from any of the three repositories, state your own board shape — five groups of five, themed weeks, harder tiles — and test it the way Example 1 tested this one: enough boards, and the right words. 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 artifacts for you or your coding agent to continue from: