I though that was what the article was talking about as well, but I think it's actually the other way around - merge the last feature branch (the one that depends on the others) into the second to last, and keep collapsing down until you've got a PR targeting master with all the changes. That way, each individual changeset can be reviewed more easily, but the whole thing is still one "unit" that will be merged into master at once. (I think that's the gist.)