How to assemble an eval kit and why? Return JSON {"steps": [list of 2+ steps], "why": "short"}.
Eval set is a set of representative examples with the expected result, on which you run the system before each change. Without it, improving the prompt is a fortune-telling: fixed one case, quietly broke five. How to collect: take real (or close to real) inputs that cover typical and edge cases; record the correct or acceptable answer for them; run the entire set before and after each edit and compare the completion rate. Even 20-50 good examples radically change the development - regressions are visible. Rule: without a set for evaluation, you do not improve the system, but only move errors from place to place.