Install any skill in seconds. Free to start, no credit card required.
Get Started Free →面向 Nature、Nature Communications 及高影响力期刊的学术语言重构与润色技能。用于中译英、英文精修、摘要压缩、段落逻辑修复、术语统一、语气校准、过度声称检查和去除模板化 AI 文风,同时严格保持数据、引用意图和科学含义不变。触发场景包括 Nature 润色、SCI 英文润色、academic editing、paper polishing、proofreading、rewrite、中文论文英译、abstract polishing、语言降重和学术表达优化。
.claude/skills/leonchaox-qinyan-nature-polishing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 24% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 16% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 18% | 0% |
把润色分成“科学含义保护、段落逻辑修复、句子精修、最终核验”四层。不得用流畅度掩盖证据缺口。
qinyan-nature-writing。python scripts/style_audit.py <file>,人工判断每条命中,不机械替换。需要完整编辑协议时读取 references/editing-protocol.md。需要按章节处理时读取 references/section-style.md。
clean:只返回可粘贴版本。annotated:返回润色稿、关键修改与科学风险。parallel:返回原文与改写的逐段对照。minimal:只修正错误和明显不自然表达,尽量保留作者声音。默认使用 annotated;用户明确要求“只给结果”时使用 clean。
text润色设定 - 章节 / 目标读者 / 模式: - 不可变事实: 润色稿 [revised text] 关键编辑说明 - [结构、语气、术语或压缩] 科学含义风险 - [原文存在歧义或证据不足之处] AUTHOR_INPUT_NEEDED - [必须由作者确认的事实]
significant/significantly。| 任务 | 读取 | |---|---| | 忠实度账本、四层编辑、语气校准、中译英 | references/editing-protocol.md | | 标题、摘要、引言、结果、方法、讨论、结论 | references/section-style.md |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 19,169 | 5,503 | -71% | 1 | 1 | 0% | 2,880 | 1,757 | -39% | 0 | 0 | — |
case-02 | fail→fail | 10,942 | 6,479 | -41% | 1 | 1 | 0% | 1,597 | 1,897 | +19% | 0 | 0 | — |
case-03 | fail→pass | 12,955 | 11,156 | -14% | 1 | 1 | 0% | 2,018 | 2,493 | +24% | 0 | 0 | — |
case-04 | fail→pass | 20,714 | 8,619 | -58% | 1 | 1 | 0% | 3,865 | 2,130 | -45% | 0 | 0 | — |
case-05 | pass→pass | 6,842 | 7,636 | +12% | 1 | 1 | 0% | 1,066 | 2,007 | +88% | 0 | 0 | — |
case-06 | pass→pass | 6,148 | 5,609 | -9% | 1 | 1 | 0% | 1,044 | 1,798 | +72% | 0 | 0 | — |
case-07 | pass→pass | 7,670 | 9,676 | +26% | 1 | 1 | 0% | 1,320 | 2,489 | +89% | 0 | 0 | — |
case-08 | fail→fail | 2,465 | 3,038 | +23% | 1 | 1 | 0% | 280 | 1,363 | +387% | 0 | 0 | — |
case-09 | pass→pass | 7,471 | 7,529 | +1% | 1 | 1 | 0% | 1,261 | 2,017 | +60% | 0 | 0 | — |
case-10 | pass→pass | 6,976 | 8,769 | +26% | 1 | 1 | 0% | 1,184 | 2,070 | +75% | 0 | 0 | — |
case-11 | pass→pass | 8,123 | 7,494 | -8% | 1 | 1 | 0% | 1,417 | 2,023 | +43% | 0 | 0 | — |
case-12 | pass→pass | 7,427 | 8,251 | +11% | 1 | 1 | 0% | 1,370 | 2,418 | +76% | 0 | 0 | — |
case-13 | pass→pass | 5,518 | 7,661 | +39% | 1 | 1 | 0% | 996 | 2,215 | +122% | 0 | 0 | — |
case-14 | fail→pass | 8,654 | 5,230 | -40% | 1 | 1 | 0% | 1,502 | 1,740 | +16% | 0 | 0 | — |
case-15 | fail→pass | 20,398 | 14,307 | -30% | 1 | 1 | 0% | 3,377 | 3,387 | +0% | 0 | 0 | — |
case-16 | fail→pass | 13,368 | 11,431 | -14% | 1 | 1 | 0% | 2,327 | 2,748 | +18% | 0 | 0 | — |
case-17 | pass→pass | 5,336 | 6,687 | +25% | 1 | 1 | 0% | 885 | 1,992 | +125% | 0 | 0 | — |
case-18 | pass→pass | 6,770 | 7,202 | +6% | 1 | 1 | 0% | 1,167 | 1,972 | +69% | 0 | 0 | — |
case-19 | pass→pass | 10,571 | 8,528 | -19% | 1 | 1 | 0% | 1,779 | 2,241 | +26% | 0 | 0 | — |
case-20 | fail→pass | 7,122 | 4,084 | -43% | 1 | 1 | 0% | 1,169 | 1,678 | +44% | 0 | 0 | — |
case-21 | fail→fail | 6,500 | 5,205 | -20% | 1 | 1 | 0% | 1,041 | 1,666 | +60% | 0 | 0 | — |
case-22 | pass→pass | 9,372 | 8,912 | -5% | 1 | 1 | 0% | 1,484 | 2,331 | +57% | 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. 22 cases were attempted. The headline lift of +27 percentage points is the difference between those two pass rates over the 22 comparable cases. 1 case got worse with the skill loaded, and it is 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.