Repository navigation
Home
English · 中文
OpenVideoHarness lets coding agents such as Claude Code and Codex make videos by writing programs, and makes the results dependable. It is a set of docs, templates and one command-line tool (bin/vh) that sits on top of rendering engines: HyperFrames, canvas-film with its Three.js layer, p5.brush, Manim and Blender.
Who it's for: people who already work with a coding agent and want explainers, science shorts, launch films, music videos, data stories or paper videos without animating by hand. Also researchers who study how agents make video, and anyone who wants to add a video type, a style or a case study.
The repository holds the rules and the reference material. This wiki is the map: where to start, how the parts fit, what works today and what comes next.
- One command installs the repo, the bundled hand-drawn engine and the agent skill.
bin/vh doctorchecks your tools. - Open Claude Code or Codex in the repo and describe the video in a sentence.
- The agent picks one of 9 video types (09 is experimental) and reads that type's workflow: engine, steps, taste, bans.
- Gate ①: two or three concepts, each a one-line idea with one frame. The look and the hook follow from the one you pick; then a one-screen outline.
-
Gate ②: a storyboard, with a keyframe for every shot, one page per section of the film. Each stop is a local review page; at the
studiolevel it is the review desk, where you tick, comment and submit. - Sound comes first. Then it writes the film scene by scene, looks at its own frames, measures the mix and fixes whatever fails a 20-item checklist. A reviewer who didn't make the film scores the draft on 8 points, the first being whether the idea holds up.
- Gate ③: the first draft, plus the 2–3 things it likes least. Once you approve, it renders the final and writes what it learned back into the docs.
Want a quick try without the stops? Ask for a "quick draft" (effort level quick). Want a say in more than the three gates, such as the hook, the main character or the theme melody? Tell the agent which decisions are yours: that is director mode.
| You are… | Read |
|---|---|
| New to it | Getting Started → FAQ |
| Making a video | Video Types → Style Library → Sound and Voice → Showcase |
| Contributing | Contributing → Roadmap → Project Status |
| Researching agent video | How It Works → research notes → research mechanisms → case studies |
Stuck? See Troubleshooting. Unfamiliar word? See the Glossary.
| Version | v0.2.1 (release, 2026-09-30). main has moved well past it without a new release: everything merged since is under Unreleased in the CHANGELOG
|
| Works today | 9 video types (09 experimental) · 32 styles · 71 shot recipes, 24 shot-grammar cards and 5 motion components · all 65 bin/vh commands, among them edit lists, archive footage, generated pictures, a sound library, a project's cast, 3D in a Three.js layer, motion blur and transparent renders, the review desk, and an asset ledger with a budget for every paid call · CI on Linux and macOS. The README's What's in the box lists the tools; Project Status says what each has been tested on |
| Next | trying the experimental parts on the real thing (type 09 on real footage, a first generated clip) · paid generation of 3D models and video clips is on hold · see Roadmap |
Details: Roadmap · Project Status
English · 中文
OpenVideoHarness 让 Claude Code、Codex 这类 coding agent 用写程序的方式做视频，而且做得稳。它是一套文档、模板和一个命令行工具 bin/vh，架在 HyperFrames、canvas-film（可加 Three.js 层）、p5.brush、Manim、Blender 这些渲染引擎之上。
适合谁：已经在用 coding agent，想做科普讲解、知识短视频、发布片、MV、数据故事或论文视频，又不想手工做动画的人；研究 agent 怎么做视频的人；想给它加视频类型、风格或案例的贡献者。
规则和参考资料以仓库为准。这个 wiki 是一张地图：从哪里开始、各部分怎么配合、现在能用什么、接下来做什么。
- 一条命令装好仓库、自带的手绘引擎和 agent skill，
bin/vh doctor检查环境。 - 在仓库里打开 Claude Code 或 Codex，用一句话说你要什么视频。
- agent 判断它属于 9 类视频（09 是实验性的）中的哪一类，去读那一类的工作流：用什么引擎、分几步、什么算好看、什么不许做。
- 关卡 ①：两三个立意，每个是一句话的点子，配一帧画面；风格和钩子跟着你选的那个走，再给一屏长的大纲。
- 关卡 ②：分镜，每个镜头一张关键帧，按片子的段落一段一页。每次停下来都是一张本地审阅页；精品档换成审阅台，你在页面里逐条勾选、点评、提交。
- 先做声音，再逐段写代码。每写完一段就看自己渲出来的帧、测混音，照 20 条清单改到合格；整片再交给一个没参与制作的 reviewer 按 8 项打分，第一项就是点子立不立得住。
- 关卡 ③：初版，外加它自己最不满意的两三处。你点头后它出成片，并把踩过的坑写回文档。
只想快速试一版？说"快速出一版"（努力程度 quick），它就不停下来问你。想在三道关卡之外多参与几件事，比如钩子、主角、主旋律？告诉它哪几件由你来定，这就是导演模式。
| 你是 | 先读 |
|---|---|
| 第一次用 | 快速开始 → 常见问题 |
| 要做视频 | 视频类型 → 风格库 → 声音与配音 → 样片 |
| 想参与贡献 | 参与贡献 → 路线图 → 项目状态 |
| 做相关研究 | 工作原理 → 研究笔记 → 学术工作里的机制 → 案例库 |
| 版本 | v0.2.1（Release，2026-09-30）。main 已经往前走了很远，只是还没发新版：之后合并的内容都记在 CHANGELOG 的 Unreleased 下 |
| 现在能用 | 9 类视频（09 实验性）· 32 种风格 · 71 张镜头配方、24 张镜头语法卡、5 个动效组件 · 全部 65 条 bin/vh 命令，其中有剪辑表、档案素材、生成图片、声音库、项目的角色设定（cast）、Three.js 层里的 3D、运动模糊和透明底渲染、审阅台，以及每次付费调用都要过预算的素材台账 · Linux 和 macOS 上的 CI。工具清单见 README 的工具一览，各自测到什么程度见项目状态
|
| 接下来 | 把实验性的部分放到真实场景里试（09 类用真实素材、第一次真的生成片段）· 付费生成 3D 模型和视频片段先放着 · 详见路线图 |