“For controlling model tone, OpenAI suggests concrete instructions about writing behavior are more effective than abstract terms like 'friendly' or 'empathetic'.”