The cards would be much better if he reversed the order of his pipeline. Generative models are easier to train and yield better results if they're conditioned on additional information (just see all the conditional GANs for images/videos). His pipeline tries to do the most challenging task first, the Description, but it would work much better if it was done last. I'd probably generate information in this order with later steps using information from all prior steps as input:

Rarity -> Colors and Mana Cost -> Type(s) -> Name -> Power/Toughness (if needed) -> Description

This approach would ensure consistency and avoid the challenge of finding the right mana cost that he mentions. Anyway, this is a super fun project and a nice write-up!