Label datasets, refine prompts, audit output for bias & accuracy