Ask HN: How do you enhance LLM app response quality?

Which aspect consumes most of your time to improve the quality of LLM responses? And how do you measure the quality?

FYI, I'm an indie developer trying to build an open-source project that helps A/B test LLM-based applications.