Given that most available models are instruction tuned now (they directly respond to instructions rather than just predicting text), not much "prompt engineering" is necessary anymore. It's remarkable how quickly this has been forgotten, the article doesn't even mention it.