本 prompt 经多轮 LLM 实测调优，综合表现首推 deepseek-v4 / v4.1-flash（no pro，over fitting）。其余模型在角色扮演上或多或少存在退化，语言风格亦常难尽如人意，Claude 则受制于偏严的安全过滤器。

This prompt has been tuned through multiple rounds of LLM testing, with deepseek-v4 / v4.1-flash (no pro, over fitting) recommended for overall performance. Other models tend to degrade in role-play to varying degrees and often fall short in language style; and Claude is constrained by an overly strict safety filter.