There is a persistent belief that getting faster at Sudoku means doing the same things more quickly. It does not. Solve time is dominated by search — the seconds spent scanning a grid looking for a deduction that is already sitting there — and search time collapses when a pattern becomes recognisable rather than derivable. The fast solver is not moving faster. They are looking less.
That reframing has practical consequences, because recognition responds to a specific kind of training and is almost unaffected by simply doing more puzzles. This guide covers what actually moves the number.
Where the time actually goes
Time yourself on a medium puzzle and pay attention to what you are doing. You will find roughly three activities:
- Placing digits. Fast, mechanical, and a small fraction of the total.
- Scanning for the next deduction. The overwhelming majority of the time, and almost all of it spent on cells that turn out to contain nothing.
- Stalling. The dead periods where you sweep the same regions repeatedly, finding nothing, because the available move is a pattern you cannot see.
Speed work that targets the first category — writing faster, tapping faster — is optimising the wrong thing. Everything below targets the second and third.
Fix your scan order and never vary it
Unstructured scanning re-covers ground you have already checked and skips ground you have not, because there is no record of where you have been. A fixed routine eliminates both problems, and after a few dozen puzzles it runs without conscious effort, freeing attention for actual pattern recognition.
A serviceable default:
- Cross-hatch digits 1 through 9 through every box.
- After each placement, follow the cascade in that digit's row, column and box to exhaustion.
- When boxes stop producing, cross-hatch by row, then by column.
- Only then pencil-mark, and look for subsets and locked candidates.
The specific order matters much less than having one. What you are building is the ability to know, without thinking, that a region has been checked — which is what stops the repeated fruitless sweeps that make up most stall time. The scanning guide goes into the mechanics.
Train recognition, not derivation
There are two ways to find a Hidden Pair. You can derive it — tally each digit's possible cells in the unit, look for two digits sharing two cells, verify. That takes maybe twenty seconds and full attention. Or you can recognise it — the configuration simply looks like a Hidden Pair, the way a familiar word looks like a word rather than a sequence of letters. That takes under a second and almost no attention.
Both routes reach the same answer, which is why solvers assume they are equivalent and that speed will arrive with experience. It does, but extremely slowly, because random puzzles present each pattern rarely and without labels. Recognition consolidates from massed, labelled, feedback-rich exposure: many instances of the same pattern close together in time, each confirmed correct or incorrect immediately.
This is the standard structure of deliberate practice in every skill domain that has been studied, and it is why chess players drill tactical motifs from position sets rather than just playing more games. Twenty puzzles that all contain Hidden Pairs will do more for your Hidden Pair recognition than two hundred random puzzles, and take a tenth of the time.
The practice pages exist for exactly this: every puzzle in a set is verified by a step-by-step solver to require the named technique, so the exposure is guaranteed rather than hoped for.
Find your actual bottleneck
General practice improves everything a little. Targeted practice improves the one thing that is costing you minutes. The trick is identifying it honestly, and the diagnostic is simple: the next time you stall for more than a minute, do not push through — stop and find out what the available move was. Use a solver, or the puzzle's walkthrough, and name the technique.
Do this five or six times and a pattern appears with uncomfortable clarity. Most solvers discover they have exactly one missing rung, and it is usually lower on the ladder than their self-image suggests. The most common by a wide margin is Omission — locked candidates — which is unglamorous, appears constantly, and is skipped by most self-taught solvers on their way to hunting for X-Wings. The second most common is failing to follow cascades, which is not a technique gap at all.
Once you know the rung, drill it in isolation until it is boring, then return to mixed puzzles. Repeat. This is a slower-sounding and much faster-working approach than solving another hundred puzzles at your current level.
Prefer cheap techniques, in cost order
Advanced solvers are not distinguished by reaching for advanced techniques. They are distinguished by reaching for them last. Every technique has a cost in search effort and a probability of being available, and the efficient policy is to run them roughly in order of cost divided by yield:
- Naked and Hidden Singles — nearly free, available constantly. Always first.
- Omission — cheap, extremely common on medium and hard puzzles. The best return in the game.
- Naked and Hidden Pairs — cheap once marked up, common.
- Triplets — moderate cost, moderately common.
- X-Wing — expensive, requires a grid-wide single-digit scan.
- Swordfish and quads — expensive and rare. Genuine last resorts.
The common failure is jumping to X-Wing the moment a puzzle feels hard, spending ninety seconds on a grid-wide single-digit scan, and finding nothing — while a locked candidate sat available the whole time. Discipline about ordering is worth more than any individual technique.
Mark up less, but mark up properly
Full pencil marks are slow to create and expensive to maintain, since every placement invalidates marks across twenty cells. On easy and medium puzzles the annotation costs more than it saves. But partial, half-maintained marks are the worst option of all, because a stale candidate is not a neutral omission — it is false information you will eventually trust.
The workable policy is binary: solve unmarked as far as scanning takes you, and when you genuinely need candidate sets, mark completely and keep them current. Never live in between. See the pencil-marks guide for notation that stays cheap to maintain.
Use the timer as a diagnostic, not a scoreboard
Timing every puzzle and watching a personal best is mildly motivating and almost useless as feedback, because a single number cannot tell you where the time went. Worse, chasing the number directly encourages the one thing that destroys accuracy: guessing on two-candidate cells to save the search.
The useful version is to time segments. Note roughly when you stall and for how long. Three forty-second stalls in one puzzle is two minutes of pure search failure, and it tells you precisely where to aim your practice. Total time tells you nothing except whether today went well.
A realistic four-week plan
- Week one — fluency in singles. Nothing but Naked Single and Hidden Single sets, unmarked, with a fixed scan order. Goal: clear an easy grid without pencil marks. This is the substrate everything else runs on and is worth more time than feels reasonable.
- Week two — locked candidates. Drill Omission in both directions, deliberately forcing yourself to check columns as carefully as rows.
- Week three — subsets. Naked and Hidden Pairs, then Triplets. Read why subsets work so the pattern is reasoning rather than a memorised shape, which is what makes it transfer to the ragged real-board versions.
- Week four — mixed and fish. Move to technique combinations, where two advanced patterns are both required, then add X-Wing. Test on the hardest puzzles.
Speed follows from this almost as a side effect. You will not have trained yourself to move faster; you will have converted a series of slow derivations into fast recognitions, and the clock will reflect it without you having hurried once.