CASE EVIDENCE / A RECORD
hylarucoder 公开的智能体工作流案例，来源为 公开代码库，复核于 2026-06-26T12:30:20Z。
原始记录：hylarucoder AI味去除: Chinese AI Text Humanization on Gemini 2.5 Pro
Chinese Brief
hylarucoder 公开的智能体工作流案例，来源为 公开代码库，复核于 2026-06-26T12:30:20Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
任务
Use Gemini 2.5 Pro's reasoning capability to detect and rewrite AI-generated Chinese text patterns ('AI味') to produce human-sounding content. The project provides a structured prompt that instructs the model to rewrite articles while preserving meaning, reducing AI detection scores from ~70% to ~17% as measured by 朱雀大模型 detector.
公开材料提供公开代码、README 或项目配置，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：Open-source prompt methodology (1,067 GitHub stars) validated specifically on Gemini 2.5 Pro. Achieves: 1000→2000 word expansion with <22% AI detection increase; 5000-word rewrite reducing AI detection from ~70% to ~17%…
Gemini 2.5 Pro is the only model validated for this workflow — the project explicitly states '仅在 Gemini 2.5 Pro 上测试通过' (only tested on Gemini 2.5 Pro) and requires a reasoning model. The model's Chinese language fluency, long-context understanding, and reasoning ability are essential for the nuanced rewriting task.
Prompt-based project, not a deployed product. Effectiveness depends on detection tool used (朱雀大模型). Author notes results may vary.