They mention a framework for protein design. But I guess you'd start out with a know sequence/structure and mutate it towards what you want.

Yes, a common way to design new structures is to 'thread' the sequence onto existing folds.

Rarely, there have been attempts to engineer entirely new folds, although it's not clear how necessary that is.

Early evidence seems to show that AlphaFold has trouble with single point mutations that change a protein's shape, so completely de-novo proteins will be a challenge:

https://www.nature.com/articles/s41594-021-00714-2

I would not expect any program to reliably predict the effects of single mutations that massively change the protein's shape unless there was enough high quality structural data and sequence data for both substates and enough signal to predict which substate the protein would adopt after mutation.

Fortunately, evolution already encoded robustness against this sort of problem into proteins and the vast majority of single point mutations are tolerated (the resulting enzymes are often nearly as active and stable as the originals).

The point of AlphaFold not being able to predict the impact of small changes is probably more about having only seen the "right" way those protein sequences usually look. Many point mutations will result in a fitness reduction and therefore the difference would not make it into the training set.

No, this is absolutely not true. For example scientists have made hundreds of point mutations and probed them, even if the "fitness" was lower (not clear what that even means in this context).

Don't forget that you can replace the three residues in the serine protease catalytic triad and still see significant proteolytic activity. Everything we know about protein activity is wrong.