Fine-tuning on simulated data outperforms prompting for agent tone of voice