How do you systematically evaluate and iterate on a prompt rather than eyeballing outputs?

Build a labeled test set of representative and edge-case inputs with expected outputs or acceptance criteria, then score prompt versions against it instead of judging one output by eye. Choose metrics fit for the task: exact/regex match or JSON-schema validity for structured tasks; reference-based metrics or task-specific checks otherwise; and an LLM-as-judge with a clear rubric for open-ended quality (validated against some human labels, since judges have biases). Track pass rate, failure categories, latency, and cost per version, and change one variable at a time so you can attribute improvements. Guard against overfitting by holding out a test split and re-running when the underlying model changes, since prompts are model-specific. Version prompts and log real production failures back into the eval set to close the loop.