Telling a model what not to do often plants the idea. Why negative instructions misfire, and five rewrites that work better.
