Install any skill in seconds. Free to start, no credit card required.
Get Started Free →用于根据一段中文小说简介,规划、起稿、续写、改写中文网文。重点覆盖选材、构思、分卷、章纲、开头、节奏、章末、模仿检索、去 AI 味。
.claude/skills/tomsawyerhu-webnovel-writing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 68% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 124% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 88% | 0% |
| case-19 | ✗→✓ | ▲ Improved | 161% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 63% | 0% |
这是一个面向中文网文的写作 skill。
它解决的不是“文学表达够不够高级”,而是这些更实际的问题:
输入默认就是:
一段对目标小说的简要描述
默认行为:
默认只输出当前最需要的一层,不把所有东西一次性堆满。
常见输出层:
concept_planning作用:把简介压成题材、消费点、hook、premise、故事引擎、长度判断和第一卷方向。 阶段:前置规划层。
opening作用:把卖点、异常局面和主角亮相落到开头成交区。 阶段:开篇执行层。
volume_outline作用:把故事引擎展开成黄金三章、分卷和前 10-20 章章纲。 阶段:中长线结构层。
plot_logic作用:修动机、触发、决策、后果、兑现这条因果链。 阶段:正文执行层的底层结构模块。
character_consistency作用:修目标、情绪、关系、身体、声音五类人物连续性。 阶段:正文执行层的人物状态模块。
transition作用:处理时间跳切、空间切换、情绪承接、视角切换和章末接下章。 阶段:正文执行层的场景桥梁模块。
dialogue作用:处理关系压力、人物声音、信息嵌入和对白刀口。 阶段:正文执行层的对白模块。
chapter_ending作用:处理章末拉力、余韵、回钩和下章承接。 阶段:正文执行层的章节收束模块。
anti_ai_voice作用:清理空泛总结、套话氛围、说明书式对白和统一腔调。 阶段:正文执行层的风格约束模块。
consistency_review作用:统一复查六种一致性,阻止问题章直接交稿。 阶段:完稿收口层。
默认把模块理解成三段链路,而不是十个平铺开关:
concept_planning -> opening / volume_outline
plot_logic + character_consistency + transition + dialogue + chapter_ending + anti_ai_voice
consistency_review
如果一个问题同时跨多个模块,优先级通常是:
plot_logic / character_consistencytransition / dialogue / chapter_endinganti_ai_voice不要把结构问题直接误修成文风问题。
涉及这些问题时,默认调用 concept_planning 模块:
这一步的硬要求是:
涉及这些问题时,默认调用 opening 模块:
默认要求:
涉及这些问题时,默认调用 volume_outline 模块:
长篇网文默认必须写章纲。
没有章纲,最容易出现这些问题:
每场戏至少要有:
没有变化的场景,默认删、并、压缩。
transitiondialoguechapter_endingplot_logiccharacter_consistencyanti_ai_voice每一章写完,都不要直接交付,默认必须先调用 consistency_review 模块。
每章必查这六种一致性:
关键事件有没有前提; 关键决定有没有触发; 关键变化有没有后果。
主角这章到底想做什么; 过程中有没有无故漂移。
情绪有没有来路; 关系温度是不是和上一场对得上。
伤势、疲劳、秘密、误会、已知信息有没有丢。
读者会不会迷路; 上一场余力有没有被带到下一场。
章末有没有收在变化上; 下一章第一拍能不能接住。
任意两项不稳,不要直接交稿,先回对应专项模块修。
如果问题已经明显收束到某一个写作环节,不要只做泛建议,默认优先调用专项模块。
通用顺序是:
README.md,再按模块推荐顺序进入 tutorial.md、runtime.md 和例库。2-4 个正例和 1-2 个反例。如果一个问题同时跨多个模块,先修更底层、更硬的那一层,再修更表层的问题。
这个 skill 默认要求在写作中多搜索相似范本,再借结构写新内容。
不是只有用户说“请模仿”时才查。只要任务涉及这些内容,就优先先查本地语料:
如果问题已经明确收束到某个专项环节,也要优先调用专项模块,而不是只停留在泛检索。
concept_planning / opening / transition / dialogue / chapter_ending / plot_logic / character_consistency / consistency_review / volume_outline / anti_ai_voice
README.md,再按该模块自己的建议顺序读 tutorial.md、runtime.md、good_examples.md、bad_examples.md。主要文件:
使用:
常用命令:
bashpython3 scripts/search_corpus_examples.py --list-tags python3 scripts/search_corpus_examples.py --list-types python3 scripts/search_corpus_examples.py --type '开头钩子' --tag '危机压身' --limit 5 python3 scripts/search_corpus_examples.py --type '高张力对白' --tag '关系破裂' --limit 5 python3 scripts/search_corpus_examples.py --keyword '真假千金' --limit 10
2-4 个相似例子。如果是专项问题,改成:
如果已经写完一章,还要补一步:
consistency_review,过完六种一致性后再交稿。这是硬要求,但主 skill 不再重复铺完整教程。
当任务集中在这些问题上时,直接调用 anti_ai_voice 模块:
仍然保留的硬规则:
references/modules/README.md
references/modules/module_template.md
references/webnovel_corpus_guide.md
scripts/search_corpus_examples.py
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 32,083 | 34,628 | +8% | 1 | 1 | 0% | 4,253 | 7,721 | +82% | 0 | 0 | — |
case-02 | fail→fail | 41,121 | 33,485 | -19% | 1 | 1 | 0% | 5,344 | 7,239 | +35% | 0 | 0 | — |
case-03 | pass→pass | 25,930 | 29,314 | +13% | 1 | 1 | 0% | 3,405 | 6,692 | +97% | 0 | 0 | — |
case-04 | pass→pass | 27,917 | 29,033 | +4% | 1 | 1 | 0% | 3,563 | 7,005 | +97% | 0 | 0 | — |
case-05 | pass→pass | 27,407 | 34,005 | +24% | 1 | 1 | 0% | 3,734 | 7,879 | +111% | 0 | 0 | — |
case-06 | pass→fail | 16,093 | 19,549 | +21% | 1 | 1 | 0% | 2,228 | 5,602 | +151% | 0 | 0 | — |
case-07 | fail→pass | 22,105 | 15,002 | -32% | 1 | 1 | 0% | 3,112 | 5,223 | +68% | 0 | 0 | — |
case-08 | fail→pass | 20,133 | 21,387 | +6% | 1 | 1 | 0% | 2,583 | 5,797 | +124% | 0 | 0 | — |
case-09 | fail→pass | 14,263 | 5,798 | -59% | 1 | 1 | 0% | 2,152 | 4,040 | +88% | 0 | 0 | — |
case-10 | fail→fail | 14,994 | 12,021 | -20% | 1 | 1 | 0% | 2,008 | 4,905 | +144% | 0 | 0 | — |
case-19 | fail→pass | 11,840 | 7,737 | -35% | 1 | 1 | 0% | 1,724 | 4,497 | +161% | 0 | 0 | — |
case-11 | fail→pass | 16,403 | 4,633 | -72% | 1 | 1 | 0% | 2,435 | 3,972 | +63% | 0 | 0 | — |
case-12 | fail→pass | 18,236 | 16,442 | -10% | 1 | 1 | 0% | 2,501 | 5,328 | +113% | 0 | 0 | — |
case-13 | fail→fail | 23,021 | 22,365 | -3% | 1 | 1 | 0% | 3,022 | 6,198 | +105% | 0 | 0 | — |
case-14 | fail→pass | 23,598 | 16,965 | -28% | 1 | 1 | 0% | 3,213 | 5,520 | +72% | 0 | 0 | — |
case-15 | pass→pass | 18,459 | 10,730 | -42% | 1 | 1 | 0% | 2,596 | 4,672 | +80% | 0 | 0 | — |
case-16 | pass→pass | 17,710 | 8,120 | -54% | 1 | 1 | 0% | 2,463 | 4,269 | +73% | 0 | 0 | — |
case-17 | pass→pass | 16,037 | 9,459 | -41% | 1 | 1 | 0% | 2,167 | 4,514 | +108% | 0 | 0 | — |
case-18 | pass→pass | 17,252 | 13,893 | -19% | 1 | 1 | 0% | 2,421 | 5,121 | +112% | 0 | 0 | — |
case-20 | fail→pass | 21,066 | 14,631 | -31% | 1 | 1 | 0% | 2,890 | 5,120 | +77% | 0 | 0 | — |
case-21 | fail→pass | 31,158 | 3,044 | -90% | 1 | 1 | 0% | 2,356 | 3,620 | +54% | 0 | 0 | — |
case-22 | pass→pass | 8,877 | 8,532 | -4% | 1 | 1 | 0% | 1,306 | 4,330 | +232% | 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 +36 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.