# Talking-Head-Recut Skill
**作者**: HeyGen
**日期**: 2026-07-11T17:40:15.000Z
**来源**: [https://x.com/HeyGen/status/2075998619325604238](https://x.com/HeyGen/status/2075998619325604238)
---
## Day 6 of 30: Sync motion graphics on any talking head video
/talking-head-recut takes a talking-head video and layers designed graphic cards on top of it, synced to what is being said: kinetic titles, stat count-ups, model cards, lower thirds, quotes. The clip plays untouched underneath. The agent reads the transcript, decides what deserves a graphic, and designs every card in your design system.
The Day 6 twist: there was no footage. We used the HeyGen CLI to access our avatar library and picked our avatar. The agent researched the story, wrote the script, generated the lip-synced presenter from the terminal, then recut its own footage with motion graphics.
## One prompt, one avatar:
We gave it a runtime, a topic to research, and asked it to use one of our avatars.
The topic: OpenAI's GPT-5.6 announcement and its three models, Sol, Terra, and Luna.
prompt:
```
We are going to create a HyperFrames video using the /talking-head-recut
skill and our avatar Tiff Tech, via the HeyGen CLI.
Create a 45-second explainer on the ChatGPT-5.6 announcement.
Specifically we should be talking about Sol, Terra, and Luna,
the three models in the 5.6 update:
1. Scan the GPT-5.6 announcement links
2. Pull the top 10 trending X posts about it
3. Use that research to write a script for the video
4. Generate the avatar A-roll with the HeyGen CLI
5. Use /talking-head-recut to turn that script and our avatar into
a captivating video with motion graphics and a URL throughout
```
What the agent actually did
1. Researched for real. It read OpenAI's announcement posts and the top trending X reactions: Sam Altman calling it "obviously the best model we have ever produced," ARC Prize verifying 92.5% on ARC-AGI-2, the Artificial Analysis scores, the API pricing. Every number on screen is sourced, not vibes.
2. Borrowed the design system from chatgpt.com. It pulled the live CSS tokens (the true-black canvas, the gray ladder, the blue accent, the pill radii) into a FRAME.md, the Day 2 discipline. Every card looks like OpenAI shipped it.
3. Cast the presenter from the terminal. It listed our avatar library with the HeyGen CLI, loaded the Tiff Tech look, checked which render engine the look supports, and generated the lip-synced A-roll straight from the command line. When the first take ran long, it re-timed the voice to land the 45-second brief instead of cutting the script.
4. Transcribed its own A-roll and synced every card to the words. Local word-level transcription drives the timing: the 92.5% count-up lands as she says the number, the ChatGPT / Codex / API graphics pop as she names each one, and the Sun, Earth, and Moon orbs land on the closing line.
5. Finished the film. Split-screen choreography that slides the avatar to a half-frame while the model cards build, a persistent openai.com URL bug, our OpenAI logo animation woven in as the end card, rendered to MP4.
Output: the avatar, the research, the script, and every graphic came from prompts. Our revisions were sentences, not paragraphs: move the rollout graphics to her left third, let the logo animation close the video alone. The agent re-cut and re-rendered each time.
## Two changes: captions / music
We made two changes. Add captions to the bottom of the video and realign the graphics so they live above them, and add music so it isn’t just her voice. The agent built 18 phrase-level captions from its own transcript timings, moved the hook title up, relocated the URL bug to the top corner, added the track I dropped and re-rendered.
Final output:
Everything you just watched came out of HyperFrames and the HeyGen CLI: the presenter, the voiceover, the research, the motion graphics, the design system, the captions. We never opened an editor, and we never pointed a camera.
Enjoying the series?
⭐ HyperFrames on GitHub: https://github.com/heygen-com/hyperframes
Github to HyperFrames HeyGen CLI: https://github.com/heygen-com/heygen-cli
Github to HyperFrames Talking-Head-Recut Skill: https://github.com/heygen-com/hyperframes/tree/main/skills/talking-head-recut
## 相关链接
- [HeyGen](https://x.com/HeyGen)
- [@HeyGen](https://x.com/HeyGen)
- [8.1K](https://x.com/HeyGen/status/2075998619325604238/analytics)
- [openai.com](http://openai.com/)
- [https://github.com/heygen-com/hyperframes](https://github.com/heygen-com/hyperframes)
- [https://github.com/heygen-com/heygen-cli](https://github.com/heygen-com/heygen-cli)
- [https://github.com/heygen-com/hyperframes/tree/main/skills/talking-head-recut](https://github.com/heygen-com/hyperframes/tree/main/skills/talking-head-recut)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [1:40 AM · Jul 12, 2026](https://x.com/HeyGen/status/2075998619325604238)
- [8,132 Views](https://x.com/HeyGen/status/2075998619325604238/analytics)
- [View quotes](https://x.com/HeyGen/status/2075998619325604238/quotes)
---
*导出时间: 2026/7/12 08:47:02*
---
## 中文翻译
# Talking-Head-Recut 技能
**作者**: HeyGen
**日期**: 2026-07-11T17:40:15.000Z
**来源**: [https://x.com/HeyGen/status/2075998619325604238](https://x.com/HeyGen/status/2075998619325604238)
---
## 30 天挑战第 6 天：在任意口播视频上同步动态图形
/talking-head-recut 接收一段口播视频，并根据语音内容叠加设计好的图形卡片：动态标题、统计数字滚动、模型卡片、下三分屏、引言等。原始视频素材在底层原样播放。代理会阅读逐字稿，决定哪些内容值得配上图形，并按照你的设计系统设计每一张卡片。
第 6 天的转折点在于：并没有现成的素材。我们使用 HeyGen CLI 访问我们的头像库并选定了头像。代理研究了这个故事，撰写了脚本，从终端生成了唇形同步的主讲人视频，然后用动态图形对自己生成的素材进行了重新剪辑。
## 一个指令，一个头像：
我们设定了视频时长，一个研究主题，并要求它使用我们的一个头像。
主题：OpenAI 的 GPT-5.6 发布公告及其三个模型：Sol、Terra 和 Luna。
指令：
```
我们将使用 /talking-head-recut 技能和 HeyGen CLI，
结合我们的头像 Tiff Tech 来制作一个 HyperFrames 视频。
创建一段关于 ChatGPT-5.6 发布的 45 秒解说视频。
具体来说，我们应该讨论 Sol、Terra 和 Luna，
即 5.6 更新中的三个模型：
1. 扫描 GPT-5.6 发布链接
2. 拉取关于它的 X 平台上前 10 条热门帖子
3. 利用这些研究为视频撰写脚本
4. 使用 HeyGen CLI 生成头像主素材（A-roll）
5. 使用 /talking-head-recut 将该脚本和我们的头像转化为
一段带有动态图形和全程 URL 显示的精彩视频
```
代理实际做了什么：
1. 真正进行了研究。它阅读了 OpenAI 的公告帖子和 X 平台上的热门反应：Sam Altman 称其为“显然是我们生产过的最好的模型”，ARC Prize 验证了其在 ARC-AGI-2 上达到 92.5% 的成绩，以及 Artificial Analysis 的评分和 API 价格。屏幕上的每一个数字都是有来源的，而非凭空感觉。
2. 借用了 chatgpt.com 的设计系统。它提取了实时的 CSS tokens（纯黑画布、灰色阶梯、蓝色强调色、圆角半径）到一个 FRAME.md 文件中，这是第 2 天的纪律。每张卡片看起来都像是 OpenAI 官方出品。
3. 从终端选角。它使用 HeyGen CLI 列出了我们的头像库，加载了 Tiff Tech 的形象，检查了该形象支持的渲染引擎，然后直接从命令行生成了唇形同步的主素材。当第一次拍摄时长超标时，它重新调整了语速以符合 45 秒的简报，而不是删减脚本。
4. 转录了自己的主素材并将每张卡片与语音同步。基于本地单词级的转录驱动了时机：当她说到数字时，92.5% 的滚动数字正好出现；当她提到 ChatGPT / Codex / API 时，相应的图形随之弹出；当她说到结束语时，代表太阳、地球和月球的图标也随之出现。
5. 完成了影片。分屏编排将头像滑向半帧画面，同时展示模型卡片，持久的 openai.com URL 角标，以及我们 woven 进结尾卡片的 OpenAI logo 动画，最终渲染为 MP4。
输出结果：头像、研究、脚本和每一张图形都来自于指令。我们的修改意见只是句子，而不是段落：把 rollout 的图形移到她的左侧三分之一，让 logo 动画单独作为视频结尾。代理每次都重新剪辑和重新渲染。
## 两处改动：字幕 / 音乐
我们做了两处修改。在视频底部添加字幕，并重新调整图形位置使其位于字幕上方；添加音乐，使其不仅仅只有她的声音。代理根据自己转录的时机构建了 18 个短语级字幕，将钩子标题上移，将 URL 角标重新定位到顶部角落，加入了我放置的音轨并重新渲染。
最终输出：
你刚刚看到的所有内容都出自 HyperFrames 和 HeyGen CLI：主讲人、旁白、研究、动态图形、设计系统、字幕。我们从未打开过编辑器，也从未使用过摄像机。
喜欢这个系列吗？
⭐ GitHub 上的 HyperFrames: https://github.com/heygen-com/hyperframes
GitHub 上的 HyperFrames HeyGen CLI: https://github.com/heygen-com/heygen-cli
GitHub 上的 HyperFrames Talking-Head-Recut 技能: https://github.com/heygen-com/hyperframes/tree/main/skills/talking-head-recut
## 相关链接
- [HeyGen](https://x.com/HeyGen)
- [@HeyGen](https://x.com/HeyGen)
- [8.1K](https://x.com/HeyGen/status/2075998619325604238/analytics)
- [openai.com](http://openai.com/)
- [https://github.com/heygen-com/hyperframes](https://github.com/heygen-com/hyperframes)
- [https://github.com/heygen-com/heygen-cli](https://github.com/heygen-com/heygen-cli)
- [https://github.com/heygen-com/hyperframes/tree/main/skills/talking-head-recut](https://github.com/heygen-com/hyperframes/tree/main/skills/talking-head-recut)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [1:40 AM · Jul 12, 2026](https://x.com/HeyGen/status/2075998619325604238)
- [8,132 次观看](https://x.com/HeyGen/status/2075998619325604238/analytics)
- [查看引用](https://x.com/HeyGen/status/2075998619325604238/quotes)
---
*导出时间: 2026/7/12 08:47:02*