In the post, author says that the GPT-3 version it apparently worked well. Which is really is amazing to me.

The prompts really read in a way where I wouldn't have expected the "correct" output.