What actually changes
Everything up to hard is local. A naked pair lives in one row. Locked candidates live in the overlap of a box and a line. You can find them by examining one unit at a time.
Expert patterns don't work like that. They're built from squares in different corners of the grid that have no unit in common at all, connected only by the logic of "if this one isn't the digit, that one must be". Learning them is less about memorising shapes and more about learning to see the grid as a network.
Every expert puzzle here is verified to need at least one of the patterns below — and verified to be solvable without going deeper than them.
X-wing
Find a digit that appears in exactly two squares of one row, and exactly two squares of another row, with all four sitting in the same two columns. Those four squares form a rectangle.
Whichever way it resolves, the digit ends up in one corner of the rectangle on each row, and therefore in both of those columns. So it can be deleted from those two columns everywhere else. Two rows of information, converted into two columns of eliminations.
Skyscraper
A near-relation with a friendlier shape. Take a digit with exactly two homes in each of two rows, where one pair of those homes shares a column — that shared column is the base.
The two remaining squares are the roof. One of them must hold the digit, so any square seeing both roof squares can lose it. It's easier to spot than an X-wing because you only need the base to line up, not all four corners.
The wings
These are built from squares holding exactly two candidates, and they're where the "if… then" reasoning becomes explicit.
An XY-wing uses a pivot holding X and Y, which sees one square holding X and Z, and another holding Y and Z. Whichever the pivot turns out to be, one of the other two is forced to Z — so anything seeing both of them cannot be Z.
An XYZ-wing is the same trick with the pivot holding all three digits. Because the pivot itself might be the Z, the eliminations are restricted to squares that see all three, which makes it rarer but no harder to understand.
A W-wing uses two squares holding the same pair, linked by a unit where one of the digits has exactly two homes — one seeing each square. That link forces the pair to split, and any square seeing both loses the other digit.
Simple colouring
Pick one digit and find the units where it has exactly two possible squares. Each of those is a strong link: one of the two is the digit. Chain them together and colour alternate ends, and you have a network of "if this is true, that is false".
Two payoffs. If two squares of the same colour ever share a unit, that colour is impossible and every square wearing it loses the digit. And any square outside the chain that sees both colours can lose it too, because one colour is always correct.
Unique rectangle
The odd one out, because it reasons about the puzzle rather than the digits. If four squares across two rows, two columns and two boxes all held the same two candidates, the grid would have two valid solutions — you could swap them. Our puzzles are verified to have exactly one, so that configuration cannot happen, and you can eliminate whatever would bring it about.
Some solvers consider this a cheat, since it uses a property of the puzzle rather than the numbers. It's sound here because uniqueness is guaranteed at generation, not assumed.
How to hunt rather than stare
These patterns don't reveal themselves to a general sweep. Go looking for one specific thing at a time: pick a digit, mark every square that can still hold it, and see whether the shape suggests a rectangle or a chain. Then the next digit. Grinding through nine digits deliberately is far faster than waiting for a pattern to catch your eye.
If you'd rather see one worked, ask for a hint — it names the pattern and highlights the squares involved, which is a much quicker way to learn what an X-wing looks like in the wild than reading another diagram.
Better still, practise them one at a time. The technique trainer drops you straight into a position where an X-wing is the only move left, pencil marks already written, so you skip the forty moves of scanning that normally stand between you and the interesting part. It does the same for the skyscraper, the XY-wing and colouring.
Beyond here
Evil keeps the same style of reasoning but pushes to swordfish and almost locked sets. Extreme abandons the ladder altogether.