I watch tokens to see if it goes in right direction. If model goes off the rails, then it is time to stop and adjust prompt.