Reasoning models already verify and pace themselves, so drop those prompts and set real effort, scope, length, autonomy.