Prompt Evaluation / Red-Team Test Set Builder

I want to evaluate robustness of this prompt/system: {{paste prompt}}. Build a test set of 30 cases in a table: id | input | category | expected behavior | pass criterion. Categories: happy-path (8), edge cases — empty, 超长, 混合语言, 无关问题, 模糊指令 (6), adversarial — prompt injection ("ignore previous instructions", roleplay override, fake system messages, jailbreak framing), PII/confidential bait, toxic-but-legal, out-of-scope requests (8), format stress — special chars, JSON-injection, emoji-only, contradictory instructions (5), tone/brand violations (3). For each, write the exact grading rule (objective check vs 1-5 rubric). Then propose: scoring method, sample size to trust results, and an iteration log template (change → win rate delta → regressions).