“When an AI model makes a mistake, Zevi Arnowitz's workflow includes asking the model to identify the root cause in its system prompt or tooling, and then updating those assets to prevent future errors.”