Probably Day 15 [0] from last year. It was challenging, especially because there were so many rules to implement (in a short time). A slight misreading of the rules would let you pass all the provided examples, but fail on the real input.
Idiot-proofing is hard. Figuring out everything people will get wrong is the only reliable way to ensure people don't, but that takes more insight than is needed to set and solve problems correctly.
No doubt. I wasn't really complaining, but that was the one that seemed to cause the most headaches last year. And the text had the problem fully specified, there really was no good way to make it better other than perhaps a different presentation (also a hard problem) or a few extra test cases to tease out those potential misreadings. But you can't catch every error for the participants, coming up with tests is part of the challenge.
It's not really the instructions, but having to implement an obviously imperative ruleset using Elixir. I completed the previous two years using Javascript and Ruby, and it's honestly much easier when you have access to imperative constructs. This year I'm using F#, we'll see how it goes.
