Every solver has had the experience: you finish a “hard” puzzle in fifteen minutes, then stall completely on a “medium” from a different app. The labels feel arbitrary because, across publishers, they are. Difficulty is real — puzzles genuinely differ in how much they demand — but the star on the cover measures it indirectly at best. This guide explains what difficulty actually consists of, how raters try to quantify it, and why this site sidesteps labels entirely.
First, what difficulty is not: clue count
The most intuitive proxy — fewer givens, harder puzzle — is nearly useless. The proven minimum is 17 clues (see the mathematics guide for that story), yet many 17-clue puzzles solve pleasantly with singles and a pair or two, while some of the hardest puzzles ever constructed — grids with names, like the famous “AI Escargot” or the “Golden Nugget” — carry 21 to 23 givens. What matters is not how many cells are empty but what kind of reasoning the empty cells require. Ten extra givens placed where they duplicate information the grid already implies add nothing; one given withheld from a critical region can force a leap in technique.
What difficulty actually is
Human difficulty decomposes into a few measurable ingredients:
- The hardest technique required. The single biggest factor. A puzzle needing only singles is easy for anyone fluent in them; a puzzle that cannot be finished without a Swordfish is hard for everyone who hasn't trained one.
- How many hard steps, and where. One X-Wing that unlocks a cascade of singles feels satisfying; three separate advanced steps with dry spells between them feels brutal. Raters count the difficult moves, not just the peak.
- The width of the search. When your next deduction is a Naked Pair, how many candidate patterns must your eyes sweep before finding it? A pattern in a crowded grid region is objectively harder to spot than the same pattern in a sparse one — difficulty is perceptual as well as logical.
- Bottlenecks. Some puzzles have exactly one productive move at a given moment. Miss it, and everything else is wasted motion. Puzzles with several concurrent openings are more forgiving at the same technique level.
How automatic raters work
Serious rating systems all share one architecture: solve the puzzle with a catalog of human techniques, ordered from cheapest to most expensive, and score what was needed. The rater repeatedly applies the simplest technique that makes progress; whenever it must reach higher in the catalog, the puzzle's score rises. The best-known example is Sudoku Explainer, whose numeric ratings (easy puzzles around 1–2, the hardest known grids above 11) became the community's common currency for comparing extreme puzzles. Commercial apps run the same design with coarser buckets: needs singles only → “easy”; adds subsets and Omission → “medium”; adds fish and beyond → “hard” or “expert.”
Two consequences follow immediately, and they explain most rating weirdness:
- Ratings depend on the catalog. If one app's solver knows Unique Rectangles and another's does not, the identical puzzle rates differently. There is no reference catalog, so there is no reference rating.
- Ratings depend on tie-breaking. When several techniques could fire, the order the solver tries them changes which path it takes — and therefore what it thinks the puzzle required. Some puzzles have an easy path and a hard path, and a mechanical rater finds only one.
Newspaper convention adds a final layer of noise: grading by average solve time panels, or by generator settings that only loosely track technique demand. “Fiendish” is marketing, not measurement.
Difficulty is relative to the solver
The deepest problem with a one-dimensional rating is that human difficulty is not one-dimensional. To a solver who has drilled fish patterns until they pop out of the grid, an X-Wing puzzle is a pleasant medium; to an equally intelligent solver who has never seen one, it is a wall. Your personal difficulty function is determined by which patterns your eyes have automated — which is why two people disagree about the same puzzle, and why your own ratings shift as you improve. A “hard” puzzle is precisely a puzzle requiring a technique you haven't internalized yet. That is not a flaw in you; it is a to-do item.
The alternative: name the techniques
Once you see ratings as compressed technique lists, the obvious move is to stop compressing. Precision Sudoku's technique picker replaces the difficulty label with the list itself: you select which techniques a puzzle should require, and the generator guarantees the puzzle needs exactly those — no unstated leap hiding at step forty, no mislabeled cakewalk. In effect, you are setting your own difficulty in the only units that mean anything: named, learnable patterns from the technique ladder.
A practical way to use this: when a rated puzzle from elsewhere defeats you, identify the step you missed (a solver or the puzzle's solution walkthrough will name it), then drill that technique here in isolation until it is boring, mixing it back into varied puzzles afterward. Difficulty, approached this way, stops being a property of puzzles and becomes a curriculum — consumed one worked example at a time.