Puzzles you can build

Four formats, and the algorithm behind each.

01
Pick a Mechanic
what a “level” is, what you need at runtime
↓
02
Write a Traversal
an algorithm that walks the network
↓
03
Generate Every Level
hundreds or thousands at once

Every word game is a data problem wearing a different costume.

Scrabble needs to answer one question: “Is this a valid word?” A flat list of terms suffices. Wordle clones need a shorter list of familiar five-letter words—and someone has to decide whether SLAVE or ABORT might trend on social media for the wrong reasons. NYT Connections needs four categories with carefully calibrated decoys—words that look like they belong together but don’t. Semantic pathfinding games need pre-computed routes through meaning-space, because you can’t brute-force hundreds of millions of permutations in real time.

Below are four formats, and the algorithm you’d write to generate each from the data. They’re patterns to adapt, not a fixed menu—your game’s mechanics may call for something none of these covers.

Download Sample Puzzles

25 complete puzzles—10 categories, 10 words each, with difficulty scores per word. Raw algorithmic output, no human editing.

↓ TSV (27 KB) ↓ JSON (136 KB)

Anagram Puzzles

For letter-tile games, anagram challenges, and Wordle-style spelling games. Scan the vocabulary for every word a player could legitimately form from a set of tiles, pick targets with enough spread between them, and attach a one-line semantic clue to each from the association data.

Puzzle Generation

LETTER SET
7 letters (configurable)
→
VOCABULARY SCAN
for each word: does it use only these letters?
→
SPELLING RULES
• letter repetition? (DEEP, LEVEL)
• smushed compounds? (SUNFLOWER)
• minimum length?
→
TARGET SELECTION
• length variety
• edit distance ≥3
• no shared stems
• no containment

Clue Generation

TARGET WORD
e.g. CASTLE
→
SEMANTIC LOOKUP
pull related words from 250M-connection network
→
MORPHOLOGICAL FILTER
reject same-stem relatives (CASTLES, CASTLED)
→
ETYMOLOGICAL FILTER
reject same-root relatives (CHÂTEAU shares Latin root)
→
SENSE DISTRIBUTION
spread across meanings if polysemous
HARD
Sample Data
Level 847 of 2,000
Letter Tiles
L I S T E N S
Target Words
LISTENS hear, audio, ears
TITLES book, movie, name
STINTS period, job, limit
LINEN fabric, sheets, cloth
NESTS home, build, tree
ISLES islands, tropical, water
Also Valid
ELITES ENLIST ENSILE INLETS ISLETS LISTEN SILENT TINSEL STIES TILES LINES LIENS INLET SITES …
Easy Medium Hard

Pathfinding Puzzles

For semantic navigation games where players traverse meaning-space—connecting concepts through chains of associations. Walk the association network to find origin–target pairs solvable in exactly 3 moves (4 hops), keep the ones with several distinct routes, and precompute hints plus backward “convergence” clues that tell players when they’re getting close.

Puzzle Generation

ORIGIN POOL
819 curated origin words
→
BRUTE-FORCE ENUMERATION
enumerate ALL 4-hop paths to every reachable candidate target
→
PATH COUNT FILTER
keep 16–43 paths (3–4 advancing choices per hop)
→
CHOKE POINT REJECTION
reject if one hop1 word dominates >60% of routes
→
SHORTCUT FILTER
reject if 2-hop or 3-hop paths exist (too close semantically)

Hint Generation

VERIFIED PATHS
only paths that actually reach target
→
FORWARD EXTRACTION
for each hop: collect words from all solutions
→
FREQUENCY RANKING
hints appearing in more paths ranked higher (more downstream options)
→
BACKWARD TRAVERSAL
from TARGET back: find all words 1–3 hops away
→
CONVERGENCE WORDS
“am I close?” clues shared across puzzles with same target

Why hints must be precalculated: Finding valid hints requires enumerating all paths—not just one—to guarantee every hint actually reaches the target. With 15 words visible at each hop, that’s 50,000+ candidate paths per puzzle. Ranking hints by how many paths use them (more paths = more downstream options) requires global knowledge. And compressing hint data for mobile delivery (74% size reduction via ID encoding) requires knowing all hints upfront. Runtime computation would add seconds of latency; precalculation makes hints instant and guaranteed correct.

MEDIUM
Sample Data
Level 89 of 779
sugar → peace
Path A
sugar → sweet → pleasant → harmony → peace
Path B
sugar → cane → plant → olive → peace
Path C
sugar → dissolve → solution → resolution → peace
Hints
sweet cane dissolve (every hint verified to reach target—no dead ends)
Converge
harmony calm treaty (1-hop from target)

Themed Categories

For NYT Connections-style games with intentional decoys—words that tempt players toward the wrong group—or for category games that need the opposite: mutually exclusive pools where no word could plausibly belong to two. You build each group from an anchor word’s strongest associations, then test every group’s neighborhood for the overlap you want (or don’t).

ANCHOR WORD
word with rich semantic profile
→
ASSOCIATION PULL
select 8 strongest associations (≤7 letters each)
→
THEMATIC ISOLATION CHECK
does this anchor’s FULL neighborhood overlap any existing output words? → reject
→
REPEAT ×8
all 8 groups must have zero semantic bleed between them
HARD
Sample Data
Level 234 of 1,500
Theme 1 Climate
monsoon carbon ozone warming ecology polar
Theme 2 Pressed
coerced ironed cider crushed mashed creased
Theme 3 Keyboard
shift return escape control space tab
Easy Medium Hard

Clue-Based Puzzles

For crossword-style games and semantic guessing games. Players identify a target word from short semantic clues—and from decoys that match some, but never all, of those clues. You pull one clue per sense of a polysemous word, filter out anything that would give the answer away morphologically, then score candidates to find decoys that land just short.

Clue Selection

ANSWER WORD
polysemous (multiple dictionary senses)
→
SENSE CATALOG
identify distinct meanings (SEAL: animal, wax, military)
→
ASSOCIATION PULL
for each sense: top associations from network
→
MORPHOLOGICAL FILTER
reject same-stem, same-root relatives (SEAL clues can’t include SEALS, SEALING)
→
4 CLUES
one strong association per sense

Decoy Selection

4 CLUES
given these 4 clues
→
CANDIDATE SCAN
score each word against all 4 clues
→
ZERO DETECTION
which clues have ZERO connection? (no network link)
→
MATCH PATTERN
record which clues each decoy matches: “0,1” or “2,3” etc.
→
SEPARATION SCORING
answer_total − best_decoy must exceed threshold

Why decoys work: Each decoy must match 2–3 of the 4 clues—enough to seem plausible—but have at least one “zero” (a clue it cannot possibly connect to). Players eliminate decoys by finding the impossible connection. The answer wins by having no zeros: it connects to all 4 clues through different senses of the word.

HARD
Sample Data
Level 512 of 3,000
Answer
army
Clues
platoon reconnaissance ambush barracks
Decoys
battalion tactics squadron
Easy Medium Hard

What the data gives you

Every format above queries the same stack—the four algorithms are just different questions asked of it. What makes them work:

The four formats above are starting points. Your game’s mechanics may call for a structure none of these covers—the data is yours to shape. See the data’s shape →

The edge cases are where all the work is—and a decade of them are already resolved in the data, so your algorithm doesn’t have to rediscover them.