Install any skill in seconds. Free to start, no credit card required.
Get Started Free →生成「水彩速写风」新闻纪实插画的提示词并出图,任何题材的文章均可使用。风格锚定:黑色墨水速写勾线 + 透明水彩薄涂 + 粉彩色板 + 现场纪实抓拍感 + 手绘场景标注(生成图中不出现画师签名、日期、时间)(参考 assets/example-watercolor-sketch.jpeg)。适用触发:1)上传或粘贴任意文章(公众号稿、新闻稿、专栏、财经/科技/文化/社会/人物报道等),要求为文章关键内容配同风格插图;2)给出一句简短要求,直接生成同风格单图;3)提到"水彩速写风""新闻速写风""水彩速写插画""watercolor sketch""reportage sketch"配图。当环境中有图像生成工具(image_generation 插件、GPT-image / gpt-image 类模型接口)时直接调用出图,没有则输出可复制的中英双语提示词。
.claude/skills/serenashenn3-art-watercolor-sketch-style/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | 44% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 3% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 13% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 37% | 0% |
为任何题材的文章或简短需求生成统一风格的水彩速写新闻纪实插画。核心资产:固定「风格锚定段」,保证同一篇文章所有插图风格统一。
参照 assets/example-watercolor-sketch.jpeg:
完整风格拆解与提示词模板见 references/style-guide.md。组装提示词前必读该文件中的锚定段原文,不要凭记忆改写。
bash python scripts/build_prompt.py --scene "<场景中文描述>" 脚本输出中英双语完整提示词。
直接用脚本组装单条提示词并出图:
bashpython scripts/build_prompt.py --scene "深夜办公室里,两名交易员隔着堆满报表的桌子争论"
image-to-url 把 assets/example-watercolor-sketch.jpeg 转成公共 URL,每张生成时以 --reference-image 传入,整套图风格即高度统一(含手写场景标注质感)。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | fail→pass | 12,051 | 9,166 | -24% | 1 | 1 | 0% | 1,833 | 2,646 | +44% | 0 | 0 | — |
case-01 | fail→fail | 14,587 | 20,916 | +43% | 1 | 1 | 0% | 2,245 | 4,777 | +113% | 0 | 0 | — |
case-02 | fail→fail | 11,730 | 10,641 | -9% | 1 | 1 | 0% | 1,988 | 3,065 | +54% | 0 | 0 | — |
case-03 | pass→pass | 10,511 | 13,579 | +29% | 1 | 1 | 0% | 1,760 | 3,534 | +101% | 0 | 0 | — |
case-04 | pass→pass | 14,045 | 25,700 | +83% | 1 | 1 | 0% | 2,075 | 3,542 | +71% | 0 | 0 | — |
case-05 | pass→pass | 11,839 | 4,079 | -66% | 1 | 1 | 0% | 1,722 | 1,854 | +8% | 0 | 0 | — |
case-07 | fail→fail | 14,239 | 12,749 | -10% | 1 | 1 | 0% | 2,356 | 3,373 | +43% | 0 | 0 | — |
case-08 | pass→pass | 5,560 | 2,523 | -55% | 1 | 1 | 0% | 914 | 1,634 | +79% | 0 | 0 | — |
case-09 | fail→pass | 11,657 | 4,573 | -61% | 1 | 1 | 0% | 1,864 | 1,917 | +3% | 0 | 0 | — |
case-10 | pass→pass | 10,318 | 5,699 | -45% | 1 | 1 | 0% | 1,634 | 2,108 | +29% | 0 | 0 | — |
case-11 | pass→pass | 12,465 | 9,343 | -25% | 1 | 1 | 0% | 2,039 | 2,711 | +33% | 0 | 0 | — |
case-17 | fail→fail | 14,405 | 8,787 | -39% | 1 | 1 | 0% | 2,220 | 2,683 | +21% | 0 | 0 | — |
case-12 | fail→pass | 11,886 | 6,400 | -46% | 1 | 1 | 0% | 1,945 | 2,208 | +14% | 0 | 0 | — |
case-13 | fail→pass | 12,617 | 5,832 | -54% | 1 | 1 | 0% | 1,947 | 2,194 | +13% | 0 | 0 | — |
case-14 | fail→pass | 12,303 | 10,430 | -15% | 1 | 1 | 0% | 2,059 | 2,827 | +37% | 0 | 0 | — |
case-15 | fail→pass | 12,361 | 4,013 | -68% | 1 | 1 | 0% | 2,248 | 1,804 | -20% | 0 | 0 | — |
case-16 | fail→pass | 10,549 | 7,157 | -32% | 1 | 1 | 0% | 1,837 | 2,404 | +31% | 0 | 0 | — |
case-18 | fail→pass | 14,912 | 6,165 | -59% | 1 | 1 | 0% | 2,304 | 2,256 | -2% | 0 | 0 | — |
case-19 | pass→pass | 13,265 | 11,338 | -15% | 1 | 1 | 0% | 2,132 | 3,159 | +48% | 0 | 0 | — |
case-20 | pass→pass | 29,145 | 19,909 | -32% | 1 | 1 | 0% | 6,174 | 5,642 | -9% | 0 | 0 | — |
case-21 | pass→fail | 8,966 | 9,153 | +2% | 1 | 1 | 0% | 1,460 | 2,698 | +85% | 0 | 0 | — |
case-22 | pass→fail | 21,104 | 17,968 | -15% | 1 | 1 | 0% | 3,470 | 4,522 | +30% | 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. 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.