Managing prompts at scale feels like herding cats. You have dozens of variations. Different versions for GPT-4, Claude, and open-source models. Some prompts work. Others fail silently. And nobody has …