Stop guessing why your LLMs break: Anthropic’s new tool shows you exactly what goes wrong