I’ll try to be clear but it’s not garanteed 
First a naming precision : I call node’s height the value that make the relief of the ground, and node’s level the elevation of a node regarding to neigbboring nodes, delimited by cliffs.
Blizzard’s approach :
Each node of the map has i’ts own level.
You can elevate one single node. Then you will have four cliffs: north-east, north-west, SE, SW. The cliffs will all have the same shape : salient.
If you want to elevate more nodes, they will be managed by group of four, snapped to pair rows and collumns of the tile map. See the left scheme above. The editor will also round the inner corners (“o” on the scheme).
The right scheme is impossible to do.
b b a a x x x x x x
b b a a x x x x x
o b b x x
b b x
The surrounding tiles will all have a cliff mesh, straight, corner or salient.
This approach has three advantages :
- every configuration works at creating a correct cliff path (chain of continuing cliffs), there are never any bug,
- straight parts of the chain have always a pair nomber of cliffs,
- diagonal parts are assemblies of the triptych “salient/corner/salient” elements.
It ensures no bug and very few different shapes for all cases : couple of straights, lone salient, diagonal triptych.
But there are tons of cliff path that are forbidden. The pair snapping is not intuitive in the editor and it may be frustrative to be unable to achieve the vision we have in mind.
My approach :
The level is stored on the tile, and given to the four connected node. When you elevate a tile, you get 9 cliffs : NE, N, NW, E, W, SW, S, SE.
But you can elevate any tile regardless of the surrounding tiles. No snapping.
This approach allows any cliff path, especially good for narrow and tortuous canyons.
But it also generates bugs. In the video, you can see red blocks when sculpting cliffs. These are detected bugs that mus be manually managed. Look at that scheme for one of the many cases :
x x x x
x x x
The empty tile needs a cliff, but the five surround elevated tiles ask for a different shape and there is no solution.
Also, in this approach, you doesn’t ensure to have grouped cliff and this lead to thin and repetitive elements. You will want to detect couple and triptych as well, but you will need more different shapes to cover every case.
Hope it helps !
Ben