That's been my experience as well. Makes sense, there's little training data available and the (public) docs and (first-party) tooling remain dreadful.