I started thinking the main challenge would be prompt design. Instead, I kept running into cases where the model output looked correct while subtly changing meaning or missing the requested tone. That became a bigger problem than generation itself.