Iterate on your LLM application with datasets, offline experiments, and the playground. Test prompt and model changes against saved inputs, and compare the results before you ship them. After you ship, use online experiments to measure the effect of a change on live traffic.