Like if you're using claude code on a feature tangential to your training pipeline it's allowed to nerf itself and damage your AI work.