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 like HyperFrames, Manim, Remotion, p5.brush 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.
- 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 on without a new release: everything merged through PR #50 is listed under Unreleased in the CHANGELOG
|
| Works today | 9 video types (09 experimental) · concept cards at gate ①, with taste rules as defaults and a short list of floors · 31 style presets to borrow from, each with a rendered 5-second swatch and its own soundtrack · three review gates, three effort levels and director mode · the spec set up front, with text-size floors by where the film is watched · 24 shot recipes and 8 full-film skeletons · playbook 00–12 · Chinese and English voiceover, bilingual captions, code-composed music, 21 built-in sound effects, level-aware mixing and audio QA · sound on showcase films 00–03 · code-driven Blender for path-traced shots (the intro film's opening and one style swatch) · CI on Linux and macOS |
| Next | a v0.3 release of what has piled up on main · every swatch reaching 8 on every review score · trying the experimental pieces on the real thing: type 09 on real footage, the rest of the Blender guide's smoke tests, a first generated clip · a person listening to the new mixes |
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、Manim、Remotion、p5.brush、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。
- 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，只是还没发新版：到 PR #50 为止合并的内容，都记在 CHANGELOG 的 Unreleased 下 |
| 现在能用 | 9 类视频（09 实验性）· 关卡 ① 的立意卡，口味规则只是默认值，底线另列一小份 · 31 个可以借用的风格预设，各带一段真渲的 5 秒样片和自己的配乐 · 三道人工关卡、三档努力程度和导演模式 · 规格一开始就定，字号下限按在哪看分档 · 24 个镜头配方和 8 条全片骨架 · playbook 00–12 · 中英配音、双语字幕、代码作曲、21 个内置音效、分层混音和音频质检 · showcase 00–03 有了声音 · 用代码驱动 Blender 渲路径追踪的镜头（介绍片的开场和一个风格样片）· Linux 和 macOS 上的 CI |
| 接下来 | 把 main 上攒下的改动发成 v0.3 · 每个样片的每一项评分都到 8 · 把实验性的部分放到真实场景里试：09 类用真实素材，Blender 指南剩下的冒烟测试，第一次真的生成片段 · 由人来听新的混音 |