“Providing a language model with the reasoning behind a constraint is more effective than using absolute negative constraints like "never do this."”