@eternityspring: Since so many people are watching, and some friends in the comments mentioned qwen3.8 27b I'll share the complete workf…

X AI KOLs Timeline Tools

Summary

该文章分享了使用qwen3.8 27b、shuohao-skills、Krea-2和MiniMax H3等AI工具进行AI短剧制作的完整工作流程,强调低成本和可自定义性。

Since so many people are watching, and some friends in the comments mentioned qwen3.8 27b I'll share the complete workflow: Use qwen3.8 27b to call shuohao-skills Generate scripts, characters, scenes, props, storyboard info, and prompts for image and video generation Then use Krea-2 to generate images Next, use MiniMax H3 to draw cards based on Krea-2's images and qwen-3.8's video prompts https://github.com/eternityspring/shuohao-skills… The entire workflow can be said to cost not a single dime, only needing a few cents for electricity. Actually, some friends researching self-deployment aren't doing it to save money, but to unleash whatever they want with the NSFW version models
Original Article
View Cached Full Text

Cached at: 09/20/26, 09:27 PM

Since so many people are watching, and some friends in the comments mentioned qwen3.8 27b
I’ll share the complete workflow:

Use qwen3.8 27b to call shuohao-skills
Generate scripts, characters, scenes, props, storyboard info, and prompts for image and video generation
Then use Krea-2 to generate images
Next, use MiniMax H3 to draw cards based on Krea-2’s images and qwen-3.8’s video prompts
https://github.com/eternityspring/shuohao-skills…

The entire workflow can be said to cost not a single dime, only needing a few cents for electricity.

Actually, some friends researching self-deployment aren’t doing it to save money, but to unleash whatever they want with the NSFW version models


eternityspring/shuohao-skills

Source: https://github.com/eternityspring/shuohao-skills

中文 English 关注作者 X

🎬 AI视频工作台

ReelBench AI短剧工作台首屏

shuohao-skills

AI 短剧制作的 skill 集合:从一本小说到直接喂生成管线的制作素材——拆角色、排大纲、出场景与道具设定、写剧本、切分镜。给 AI 编码 agent 用,Claude Code 和 codex 都能跑

整条管线长这样——改编大纲收敛结构,剧本、场景、角色三者同步迭代,分镜只做输出不做新决定

AI 短剧制作流程图
Skill做什么
novel-outline把一本小说改编成短剧大纲五件套:改编说明、人物表、爽点表、分集梗概、资产清单(含叙事道具表)。14 道质量门全部脚本检查,支持已有大纲的体检模式
novel-characters把大纲定下的角色做成角色设定集:人物画像、形象提示词、音色提示词、角色设定图。吃 outline.json 预填角色表,报告语言与出图风格可选
novel-art给 AI 短剧出美术设定集(场景 + 叙事道具):一致性锚点、光照与状态变体、尺度参照、无人无手白底提示词。吃 outline.json 预填清单,11 道质量门全部脚本检查
novel-script给 AI 短剧写剧本:场次 + 节拍流(动作与台词交替),逐集时长按语速确定性折算,钩子前 3 拍冷开场兑现是门,台词本按角色聚合带音色提示词直接对接 TTS。10 道质量门全部脚本检查
novel-storyboard给 AI 短剧出分镜:段(一次生成 ≤15 秒)→ 分镜(2–5 秒硬门)→ 分镜图(主图钉 0.00 秒、子图钉各自切点),MiniMax H3 提示词的对齐指令与切点时刻逐字对账;分镜图拿设定图当参考图真出图,export 一键出投产包。17 道质量门全部脚本检查

五个 skill 的报告都支持中英双语界面:默认中文,render --lang en 出全英文报告(数据内容保持原文)。

AI 短剧交流社群

我建了一个付费AI视频交流群,讨论 AI 视频的工作流、工具和实操。交流群 ReelBench AI 视频工作台是两项独立服务。

有兴趣的加我:微信 hao_dev,添加时备注 github

烁皓微信二维码

合成一张单页

五段的报告可以合成一张单页,左侧导航切换——有哪几段就出哪几个面板

node scripts/report.mjs --from <demo目录> --out report.html

--from 按下面的工作目录约定自动发现五份 json;也可以逐个指定(--outline --cast --art --script --storyboard)。只跑了角色那一段就只有一个面板,不报错。

它是组装器,不是独立 skill:不 import 任何 skill 的代码,而是调各自的 render --html 拿产物再拼装。所以五个 skill 一行不改、各自仍然独立可跑、可以单独拷走;某个 skill 改了渲染,这边自动跟上。

合并时处理三件事——这三件都在组装器里做,不侵入 skill

  • 样式串味。五份报告共用 57 个类名,其中 13 个同名不同定义(.copy .kpis .badge .chip……),所以给每份样式的每条选择器加作用域前缀
  • 脚本串味。各报告的脚本都是 document.querySelector('.expo') 这种全局查询,合成一页后只会命中第一个——五个导出按钮会全废。做法是给每份脚本套一层作用域代理
  • 图片路径。各报告的图相对自己那份 json 的目录(images/…E01-01/f1.png),合成后按输出文件的位置重算

默认一次显示一个面板(五份加起来将近六十万字符)。左下角「平铺全部」把所有面板同时展开,Cmd+F 恢复全局搜索。数字键 15 切面板,#pane-script 这样的深链可以直接分享到某一屏。

node scripts/report-selftest.mjs   # 92 项断言,不起浏览器

丢一本小说进去,出这五套:

novel-outline · 短剧改编大纲

短剧改编大纲报告

novel-characters · 角色设定集

角色设定集报告

novel-art · 美术设定集(场景 + 道具,设定图为 skill 实际生成)

美术设定集报告

novel-script · 剧本(时长仪表 + 分集剧本 + 台词本)

剧本报告

novel-storyboard · 分镜(分镜节奏带 + 主/子分镜图为 skill 实际生成 + H3 提示词)

分镜报告

安装

git clone https://github.com/eternityspring/shuohao-skills.git
cd shuohao-skills
./scripts/install.sh

自动检测本机装了 Claude Code 还是 codex,把所有 skill 软链过去——git pull 之后立刻生效,不用重装。

./scripts/install.sh novel-characters   # 只装某一个
./scripts/install.sh --codex            # 只装到 codex
./scripts/install.sh --uninstall        # 取消软链

不想用脚本就自己链:

ln -s "$PWD/skills/novel-characters" ~/.claude/skills/novel-characters
ln -s "$PWD/skills/novel-characters" ~/.codex/skills/novel-characters

前置条件

必需?说明
Node必需≥ 18。skill 的脚本只用标准库,没有 npm 依赖,不需要 install
模型额度必需用你当前会话的额度,不需要任何 API key
codex CLI可选出图才用得上(走内置 $imagegen)。没有就跳过出图,其余产出照常

仓库约定

每个 skill 一个目录,自包含、可以单独拷走

skills/<skill-name>/
├── SKILL.md          给 agent 读的工作流(必需)
├── README.md         给人读的说明
├── scripts/
│   ├── <name>.mjs    确定性工具,零依赖
│   └── selftest.mjs  自测,不调模型(必需)
├── references/       按需加载的详细指令
├── examples/         自带样例,同时当测试夹具
└── assets/           截图

两条硬要求:

  • 每个 skill 必须有 SKILL.md
  • 每个 skill 必须有 scripts/selftest.mjs不调用模型、不花额度,覆盖全部确定性逻辑

加新 skill 之前,先把全部自测跑一遍:

for f in skills/*/scripts/selftest.mjs; do node "$f"; done

没有配 CI——自测足够快(1 秒),本地跑一次比等 CI 更省事。只在 macOS + Node 24 上验过;代码没有平台相关调用,Linux 和更低版本 Node 理论上没问题,但没验。

端到端 demo 工作目录约定

把一本小说从头跑完五段(角色 → 大纲 → 美术 → 剧本 → 分镜),会产出大量 *.json / *.md / *-report.html不要平铺在根目录,按五个 skill 各建一个目录归档,一眼对应流水线五段:

<demo>/
├── outline/       ← novel-outline 产出:<剧>-outline.json / .md / -report.html
├── characters/    ← novel-characters 产出:<剧>-cast.json / .md / -report.html
├── art/           ← novel-art 产出:<剧>-art.json / .md / -report.html
├── script/        ← novel-script 产出:<剧>-script.json / .md / -report.html
├── storyboard/    ← novel-storyboard 产出:<剧>-storyboard.json / .md / -report.html
│   ├── manifest.json  ← export 产出
│   ├── E01-01/        ← export 的分镜投产包,每段一个文件夹(prompt.md + f1..fN.png)
│   ├── E01-02/
│   └── …
├── docs/          ← 自己写的使用说明、PR 草稿等(与机器产物解耦)
└── scripts/       ← 跑管线的辅助脚本(探索期脚本用 _ 前缀保留溯源)

约定要点:

  • 每个 skill 一个目录,装它自己的 json / md / html 三件套,加新角色/场景只往对应目录放,不污染根目录
  • 分镜的 manifest.jsonE01-0x/ 投产包一起归 storyboard/,就是 export --out storyboard 的原样产出。段文件夹不要再往下收一层(例如收进 segments/):分镜报告里的图走相对路径 <段号>/f<切序>.png,报告 html 与段文件夹必须同级,多套一层目录,报告里的图会静默全变成「未生成」占位——实测把 10 个段文件夹移进 segments/ 之后,内嵌图从 2 张变 0 张,报告不会报错
  • 报告 HTML 与生成的图/视频可由 render 重跑再生——进版本控制时建议只提交 json / md / docs / scripts,报告 HTML 和分镜 png.gitignore 排除,保持仓库轻量
  • 用法类文档(如各报告的使用说明)放 docs/,与 skill 自动生成的产物分开,方便单独维护

这套结构来自《渡口》端到端 demo 的实际归档经验,demo 的工作目录在本仓库之外,这里只固化约定。

Star 趋势

Star 趋势曲线

License

Apache 2.0

烁皓 (@eternityspring): 我说MiniMax H3能干活,好多人都不信 可事实就是如此

如果自己再部署个 krea-2 出图,相当于做视频0成本的 你的想象力,外加几毛钱的电费,大有可为的!

Similar Articles

@crypto_dazui: https://x.com/crypto_dazui/status/2053470085726314687

X AI KOLs Timeline

This article shares a complete workflow and detailed prompts for creating live-action short dramas using AI tools, covering story generation, character and scene setup, and storyboard creation. It aims to help creators master AI-powered short drama production techniques for free.

@eternityspring: Making AI short dramas—anyone can have AI write a script; but writing a script that can actually produce a good short drama? Hard! Today this gap is filled: shuohao-skills/novel-script is open-sourced. · Episode synopses generate structured scripts, with each episode designed with an opening hook and a cliffhanger ending · Each episode has a runtime limit: dialogue at …

X AI KOLs Timeline

Shuohao open-sourced the AI short drama production skill set shuohao-skills, including a four-piece toolkit for character breakdown, outline planning, art direction, and scriptwriting. It supports Claude Code and codex, open-sourcing the entire pipeline from novel to production-ready script.

@joshesye: https://x.com/joshesye/status/2052599953193566214

X AI KOLs Timeline

A detailed tutorial introducing four methods for maintaining character consistency and plot coherence when creating AI short dramas using Seedance 2.0 and GPT-image2, including extending reference videos, using keyframes as the first frame, compositing multiple video segments, and converting storyboards to video.

@iluciddreaming: AI short drama production, the workflow is already running. Toonflow, with 11.6k stars on GitHub: Throw in a script, AI writer breaks it down, intelligent storyboarding generates shot scripts, workflow runs automatically. Free and open source. The most noteworthy point: it actually works, not a demo.

X AI KOLs Timeline

Toonflow is a free and open-source AI tool that automatically breaks down scripts into shot scripts and generates storyboards, with an automated workflow that is truly usable.

@Shenmeili1213: https://x.com/Shenmeili1213/status/2065051166984351877

X AI KOLs Timeline

This article uses the case of Wang's company using AI to produce short videos and successfully monetize them. It details the AI workflow in script generation, editing, private traffic acquisition, and customer maintenance. It points out that AI can greatly reduce content production costs, and ordinary people can enter the market through three paths: niche accounts, serving offline business owners, or collaborating with partners.