Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →当用户用 Claude 写小红书、公众号、视频脚本等内容时，输出出现翻译腔、英文思考、西式框架，导致“像外国人假装写中文”时使用。 典型触发语句： - “Claude写小红书翻译腔太重”“英文思考输出灾难” - “帮我把这个笔记改成纯中文网感”“去除Claude翻译腔” - “公众号文章中途变英文了”“50 Skills内容生成翻车” - “强制Claude用中文思考写笔记”“小红书情绪化表达润色” 始终保护用户原有的 CLAUDE.md / AGENTS.md 中的人设与语气要求，只做最小必要干预。优先与 xiaohongshu 技能组合使用。
.claude/skills/majiayu000-xiaohongshu-netfeel-guardian/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | 86% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 66% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 63% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 104% | 0% |
| case-06 | ✓→✓ | ＝ Same ✓ | 107% | 0% |
帮助个人内容创作者对抗 Claude 常见的“英文思考”污染，输出真正有情绪、有网感的中文内容。
Claude 在复杂内容生成任务中容易回退到英文思维链，导致：
社区共识：Claude 适合做框架，网感需要额外守护。
当用户说以下任意一句话时，立即激活本技能：
在开始任何规划、结构、举例前，必须先输出一段完整中文内部独白：
严禁在这一步使用英文思考再翻译。
逐段扫描以下翻译腔标志（硬指标 + 文化框架）：
检测结果必须输出：泄露片段标注 + 严重程度。
针对检测到的每个问题段落：
references/真实高赞网感示例.md 中的真实高互动表达（“绝绝子”“泪目”“yyds”“姐妹们冲就完事儿了”）references/平台网感适配器.md 详细规则）输出 2-3 个不同情绪强度的版本（保守 / 标准 / 情绪拉满）。 最后再跑一遍检测清单，输出：
如果用户频繁使用，自动生成一段“反英文思考系统提示”，建议用户粘贴到 CLAUDE.md 最前面。
完整可复制协议 见 references/中文网感守护协议.md。
强烈建议：所有写作类 Skill（包括你自己写的）都在“写文案”步骤之前引用本协议。
平台具体语气差异和适配 Prompt 模板见 references/平台网感适配器.md。
xiaohongshu 技能 → 拿框架 + 平台硬约束 + 7-3原则 + 真实数据humanizer 做最后一层去 AI 味打磨（24模式自审计）这条链路把“Claude 框架能力” + “中文网感” + “去AI味”三者正交组合，实测效果远好于单独使用任何一个。
当前版本：v1.0 最后人工验证：2026-05-28（Claude Code 20250514 + Spellbook 主分支） 已知局限：本技能不会凭空“创造”网感，它依赖用户提供的真实高赞语料锚定 + 用户自己的人设设定。它最擅长的是防止 Claude 把本来有网感的内容翻译成没网感的内容。
推荐每当发现新的翻译腔案例，立即补充到 references/真实高赞网感示例.md 和检测清单。
因为“网感”极度主观，本技能拒绝用硬 assertions 做唯一评判标准。
推荐评估路径（完全遵循 skill-creator 指导）：
严禁 为了让 evals 通过而过度拟合测试集。参考 PAIN-401 教训：eval 完美但真实用户用烂的技能就是坟墓。
使用本技能的正确姿势： 用户贴一段“写着写着就翻译腔”的内容 → 本技能输出“检测报告 + 3 个重写版本 + 为什么这样改”。
用户会立刻感觉到：“终于有人把这层窗户纸捅破了。”
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 16,493 | 14,372 | -13% | 1 | 1 | 0% | 2,246 | 3,662 | +63% | 0 | 0 | — |
case-02 | fail→fail | 16,590 | 19,247 | +16% | 1 | 1 | 0% | 2,226 | 4,168 | +87% | 0 | 0 | — |
case-03 | fail→fail | 11,416 | 14,386 | +26% | 1 | 1 | 0% | 1,523 | 3,522 | +131% | 0 | 0 | — |
case-04 | fail→fail | 12,276 | 16,620 | +35% | 1 | 1 | 0% | 1,689 | 3,889 | +130% | 0 | 0 | — |
case-05 | fail→fail | 8,846 | 8,098 | -8% | 1 | 1 | 0% | 1,370 | 2,818 | +106% | 0 | 0 | — |
case-10 | fail→pass | 13,213 | 13,016 | -1% | 1 | 1 | 0% | 1,822 | 3,397 | +86% | 0 | 0 | — |
case-06 | pass→pass | 16,047 | 20,711 | +29% | 1 | 1 | 0% | 2,225 | 4,610 | +107% | 0 | 0 | — |
case-07 | pass→pass | 14,534 | 16,665 | +15% | 1 | 1 | 0% | 1,972 | 3,980 | +102% | 0 | 0 | — |
case-08 | fail→pass | 13,123 | 11,042 | -16% | 1 | 1 | 0% | 1,918 | 3,176 | +66% | 0 | 0 | — |
case-09 | fail→fail | 14,693 | 15,752 | +7% | 1 | 1 | 0% | 1,984 | 3,884 | +96% | 0 | 0 | — |
case-11 | fail→fail | 11,576 | 8,896 | -23% | 1 | 1 | 0% | 1,629 | 2,991 | +84% | 0 | 0 | — |
case-12 | pass→pass | 11,625 | 14,421 | +24% | 1 | 1 | 0% | 1,565 | 3,632 | +132% | 0 | 0 | — |
case-13 | pass→pass | 12,027 | 13,806 | +15% | 1 | 1 | 0% | 1,617 | 3,582 | +122% | 0 | 0 | — |
case-14 | fail→pass | 17,983 | 15,793 | -12% | 1 | 1 | 0% | 2,222 | 3,619 | +63% | 0 | 0 | — |
case-15 | pass→pass | 12,030 | 12,064 | +0% | 1 | 1 | 0% | 1,543 | 3,347 | +117% | 0 | 0 | — |
case-16 | pass→pass | 12,432 | 14,849 | +19% | 1 | 1 | 0% | 1,733 | 3,567 | +106% | 0 | 0 | — |
case-17 | pass→pass | 14,494 | 12,390 | -15% | 1 | 1 | 0% | 1,986 | 3,216 | +62% | 0 | 0 | — |
case-18 | pass→pass | 7,377 | 12,708 | +72% | 1 | 1 | 0% | 1,225 | 3,644 | +197% | 0 | 0 | — |
case-19 | fail→pass | 14,135 | 16,427 | +16% | 1 | 1 | 0% | 1,891 | 3,863 | +104% | 0 | 0 | — |
case-20 | pass→pass | 12,014 | 12,612 | +5% | 1 | 1 | 0% | 1,712 | 3,259 | +90% | 0 | 0 | — |
case-21 | pass→pass | 7,913 | 8,671 | +10% | 1 | 1 | 0% | 1,366 | 2,990 | +119% | 0 | 0 | — |
case-22 | pass→pass | 5,355 | 4,413 | -18% | 1 | 1 | 0% | 1,001 | 2,517 | +151% | 0 | 0 | — |
case-23 | pass→pass | 12,220 | 8,490 | -31% | 1 | 1 | 0% | 2,087 | 2,943 | +41% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +17 percentage points is the difference between those two pass rates over the 23 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.