Please Lichess, make it happen!. An endgame training from the endgame tablebase. This is exactly how chesstempo does it (I think, or at least, the tablebase plays out the moves till the end). And the puzzles would be generated or taken from actual games, but based on specific themes, for example
(R+P vs R)
(K+P vs K)
(Q+P vs K)
(Q vs K + f or c pawn)
(Q vs K + a or h pawn)
etc...
(R+P vs R)
(K+P vs K)
(Q+P vs K)
(Q vs K + f or c pawn)
(Q vs K + a or h pawn)
etc...