Really good article. I was discussing this with Shreya (the author), and an interesting insight was that prompting a LLM to follow these instructions do not work reliably.

I’ve had similar frustrations. Maybe the next thing to try is fine tuning? Curious what others think.