# Video Agent
**作者**: HeyGen
**日期**: 2026-07-23T21:38:39.000Z
**来源**: [https://x.com/HeyGen/status/2080407266755510452](https://x.com/HeyGen/status/2080407266755510452)
---
# Day 18 of 30: Using HyperFrames in Video Agent
For seventeen days, your agent has lived in your terminal. You prompt, it writes code, and a video comes out the other side. Day 18 flips the setup: Today your agent lives inside HeyGen itself.
It is called Video Agent, and it is the same idea the whole series has been built on, delegation instead of tools, delivered as a product. No CLI, no repo, no skills to install. You describe the video, and an agent plans it, scripts it, casts it, edits it, and hands you a finished cut.
## What Video Agent is
Every video tool you have ever used handed you parts: a timeline, a stock library, a caption button. Assembly was your job.
Video Agent is not a tool, it is a crew. Give it one prompt and it constructs the whole video from the ground up: the script, the avatar on camera, the voice, the scene structure, the b-roll behind the graphics, the pacing, the captions. Everything a finished video needs, decided and assembled for you.
And here is the part that matters, the part Day 17 should make feel familiar: it plans before it builds. Submit a prompt and the agent comes back with a video plan, a summary of how it intends to structure the piece, scene by scene. You read the plan, you push back in chat, and only when you say proceed does production create the video.
Open app.heygen.com and it is right on your Home page.
The setup is three decisions:
- Choose your avatar. You can use your digital twin or pick an avatar from the library.
- Voice. Your avatar's default voice attaches automatically, and you can swap it for any voice in the catalog.
- Brand System (optional). Turn it on and your colors, fonts, and logo get applied to everything the agent builds. Worth setting up once so every video after ships on-brand by default.
## How to prompt for styles
Here is how it works: describe the video you want, then add a short style paragraph. What the graphics should look like, the colors, the energy, how things move. That paragraph does most of the work, because the agent reads it and makes every visual decision through it. Anything you do not say, it decides for you.
It looks like this:
prompt:
```
Make a 45-second portrait video introducing our new app to Gen Z.
Style: Arcade Cabinet – Pixel Power-Up. CRT Black #111111, Neon Cyan #00E5FF, Electric Magenta #FF2DAA, Power-Up Yellow #FFD400, Pixel Grass Green #4CFF4C. Chunky pixel art, scanlines, glowing vector HUD, and bold bitmap typography. Sprites pop in with 8-bit bounce, score counters tick rapidly, and power-ups burst into pixel particles. Transitions are CRT glitches, screen wipes, and coin-in flashes. Playful, nostalgic, and unmistakably arcade.
```
Read that style paragraph back and notice the structure. A name for the style. The exact palette, hex codes and all. The art direction. How things move. What the transitions are. One closing line for the vibe. Five or six sentences, and every graphic in the video now has a rulebook.
The style paragraph carries the direction. And if you already have a script, paste the whole thing in and the agent follows it scene by scene, building the visuals around your exact words. Anything you attach, slides, images, logos, gets folded in too.
## The actual prompt
The arcade paragraph shows the shape. Here is the real thing: the exact prompt behind the video running through the rest of this guide, a 30-second real estate market update for Austin, Texas, presented by Ava on our elevated real estate brand system. Same structure, short ask plus style paragraph, with one more layer: the full script, scene by scene, so the agent builds around the exact words.
Two things to lock in here too. Put the duration right in the prompt, this one asks for 30 seconds, so the agent paces to it. And decide how much rope you give the agent: you can plan with it, reviewing and refining before it builds, or put it on autopilot and let it run straight through.
prompt:
```
Make a 30-second portrait real estate market update for Austin, Texas, presented by Ava. Follow the script below exactly, scene by scene. Use motion graphics for every stat, and stock footage of the Austin skyline and neighborhoods for texture between them.
Style: Elevated Estates - Quiet Luxury. Ivory #F7F4EE, Charcoal Ink #1E1E1C, Sage #A6BB94, Deep Forest #24352A, Brushed Brass #C9A227. Editorial serif headlines over generous whitespace, hairline rules, small-caps labels, and full-bleed architectural photography with soft depth. Numbers count up gently, chart lines draw themselves in, and photography drifts in slow parallax. Transitions are soft crossfades, ivory paper wipes, and hairline reveals. Calm, polished, and quietly luxurious.
Script:
Scene 1 (0-5s) - Ava on camera, skyline behind her.
"Austin, here's your market minute for July."
Scene 2 (5-12s) - Motion graphic: median price counts up to $450K, sage trend line drawing upward.
"The metro median now sits at four hundred fifty thousand, up one percent from last summer, and prices are firming."
Scene 3 (12-19s) - Motion graphic: inventory bars stepping down, brass accent on the final bar.
"Inventory tells the story. Active listings are down fifteen percent from a year ago, at four and a half months of supply."
Scene 4 (19-25s) - Stock footage of a front door and porch, stat card overlay: 24 DAYS.
"And well-priced homes are moving in about twenty-four days."
Scene 5 (25-30s) - Ava on camera, hairline rule and small-caps CTA card.
"The window favors sellers again. Thinking about a move? Let's talk before fall."
```
## Editing to make it post-perfect
The first cut is a draft, and there are two ways to close the gap between draft and post-ready.
Path one: keep talking. The chat does not end when the video renders. On this run, scene two was the holdout: the background was not right. So we told the agent, right in the same chat, to change the background of that scene. It re-cut scene two while every other scene held still exactly as approved. Here is that round-trip. If Day 17's comment-per-card flow felt right, this is the same muscle.
Path two: open a copy in AI Studio. When you want hands on the actual cut, send a copy to the editor and take full manual control: script and text, avatar scenes, voice, timing, transitions, layout, media swaps from the stock library or your uploads, music, and caption fine-tuning. The chat handled this run's fix, but when the last five percent (a caption that crowds a face, a cut that lands a beat late, a music level) needs a human hand, this is where the deepest control lives, one click away. The script sits on the left, the scene is on the canvas, and Ask AI, your avatar, media, music, and captions run down the right rail:
The workflow that comes out of this: prompt for the 90, edit for the 10. Let the agent do the assembly it is good at, then spend your taste where taste is the whole job.
## The final output
One prompt, one plan, one background changed with a sentence. This is the Austin market update that came out the other side:
Output:
## The Unlock
Seventeen days of this series taught you to direct an agent through a terminal. Video Agent is that same working relationship with the terminal removed: describe, review the plan, direct with sentences, polish the last mile yourself. The skill you have been building all month, saying what you want clearly enough that an agent can build it, turns out to be the whole interface.
⭐ If this series is useful, star the HyperFrames repo.
Try it: app.heygen.com, the Video Agent button on your Home page
Prompting guide: heygen.com/blog/video-agent-prompt-guide
Day 18 down. 12 to go.
## 相关链接
- [HeyGen](https://x.com/HeyGen)
- [@HeyGen](https://x.com/HeyGen)
- [3.1K](https://x.com/HeyGen/status/2080407266755510452/analytics)
- [app.heygen.com](https://app.heygen.com/)
- [star the HyperFrames repo](https://github.com/heygen-com/hyperframes)
- [app.heygen.com](https://app.heygen.com/)
- [heygen.com/blog/video-agent-prompt-guide](https://www.heygen.com/blog/video-agent-prompt-guide)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [5:38 AM · Jul 24, 2026](https://x.com/HeyGen/status/2080407266755510452)
- [3,118 Views](https://x.com/HeyGen/status/2080407266755510452/analytics)
- [View quotes](https://x.com/HeyGen/status/2080407266755510452/quotes)
---
*导出时间: 2026/7/24 09:37:21*
---
## 中文翻译
# Video Agent
**作者**: HeyGen
**日期**: 2026-07-23T21:38:39.000Z
**来源**: [https://x.com/HeyGen/status/2080407266755510452](https://x.com/HeyGen/status/2080407266755510452)
---
# 30 天挑战第 18 天：在 Video Agent 中使用 HyperFrames
过去的十七天里，你的 Agent 一直生活在你的终端中。你发出提示，它编写代码，一段视频便由此生成。第 18 天彻底改变了这一设定：今天，你的 Agent 直接生活在 HeyGen 内部。
它被称为 Video Agent，其核心理念贯穿了整个系列：委托而非工具，并以此作为产品交付。没有命令行界面（CLI），没有代码仓库，无需安装任何技能。你只需描述视频，一个 Agent 就会负责策划、编剧、选角、剪辑，并最终交给你一份成片。
## Video Agent 是什么
你用过的每一个视频工具提供给你的都是零散的部件：一条时间轴、一个素材库、一个字幕按钮。组装工作则由你自己完成。
Video Agent 不是一个工具，而是一个团队。只需给它一个提示，它就能从零开始构建整个视频：脚本、出镜的数字人、声音、场景结构、图形后的空镜（B-roll）、节奏、字幕。成片所需的一切，都由它为你决定并组装。
以下是关键部分，也是第 17 天的内容会让你感到熟悉的地方：它在构建之前会先进行规划。提交一个提示，Agent 会返回一个视频计划，总结它打算如何逐个场景地构建这部视频。你阅读计划，在聊天中提出反馈，只有当你点击“继续”时，制作阶段才会开始生成视频。
打开 app.heygen.com，它就在你的首页上。
设置只需做三个决定：
- 选择你的数字人。你可以使用你的数字分身，或者从库中选择一个数字人。
- 声音。你的数字人默认声音会自动附加，你可以将其换成目录中的任何声音。
- 品牌系统（可选）。开启它，你的颜色、字体和 Logo 将被应用到 Agent 构建的所有内容中。值得一次性设置好，这样之后的每个视频默认都符合品牌调性。
## 如何用提示词描述风格
它是这样运作的：描述你想要的视频，然后添加一个简短的风格段落。描述图形应该是什么样子，颜色，能量感，以及物体的运动方式。这个段落完成了大部分工作，因为 Agent 会阅读它，并据此做出每一个视觉决定。任何你没有说明的事情，它都会替你决定。
看起来像这样：
提示词：
```
制作一段 45 秒的竖屏视频，向 Z 世代介绍我们的新应用。
风格：街机柜——像素充能。CRT 黑色 #111111，霓虹青 #00E5FF，电光洋红 #FF2DAA，充能黄 #FFD400，像素草绿 #4CFF4C。厚重的像素艺术，扫描线，发光的矢量 HUD，和粗体的位图排版。精灵图以 8 位弹跳效果弹出，分数计数器快速跳动，强化道具炸裂成像素粒子。转场采用 CRT 故障、屏幕擦除和投币闪光。俏皮、怀旧，且绝对具有街机感。
```
回读那个风格段落，注意它的结构。一个风格的名称。精确的调色板，包括十六进制代码。艺术指导。物体如何运动。转场是什么。用一句话收尾定义氛围。五六个句子，视频中的每一个图形现在都有了一本规则手册。
风格段落承载了指导方向。如果你已经有了脚本，把它全部粘贴进去，Agent 会逐个场景地遵循它，围绕你确切的文字构建视觉效果。你附加的任何东西，幻灯片、图片、Logo，也会被融合进去。
## 实际提示词
街机那段展示了形式。这是真实案例：本指南其余部分播放的视频背后的确切提示词，一段 30 秒的德克萨斯州奥斯汀房地产市场更新，由 Ava 在我们的高端房地产品牌系统上展示。相同的结构，简短的需求加上风格段落，多了一层：完整的脚本，逐个场景，以便 Agent 围绕确切的文字进行构建。
这里还有两点要确定。把时长直接写在提示词里，这个要求是 30 秒，所以 Agent 会据此调整节奏。并决定你给 Agent 多大的自主权：你可以与它一起规划，在构建前审查和细化，或者开启自动驾驶让它直接运行到底。
提示词：
```
制作一段 30 秒的竖屏视频，由 Ava 展示德克萨斯州奥斯汀的房地产市场更新。严格按照以下脚本，逐个场景执行。每个统计数据使用动态图形，使用奥斯汀天际线和街区的空镜素材作为它们之间的纹理。
风格：高端地产——静奢风。象牙白 #F7F4EE，炭墨色 #1E1E1C，鼠尾草绿 #A6BB94，深森林绿 #24352A，拉丝黄铜色 #C9A227。在宽敞的留白上编辑衬线标题，细线规则，小型大写标签，以及具有柔和景深的全出血建筑摄影。数字温和向上滚动，图表线自动绘制，照片以慢视差漂移。转场采用柔和交叉淡化、象牙白纸张擦除和细线揭示。冷静、精致，且静谧奢华。
脚本：
场景 1 (0-5秒) - Ava 出镜，身后是天际线。
“Austin，这是你 7 月的市场一分钟。”
场景 2 (5-12秒) - 动态图形：中位数价格向上计数至 45 万美元，鼠尾草绿趋势线向上绘制。
“都市区中位数现在位于四十五万美元，比去年夏天上涨百分之一，价格正在走稳。”
场景 3 (12-19秒) - 动态图形：库存条状图逐级下降，最后一根条状图带有黄铜色强调。
“库存说明了问题。活跃房源比一年前下降百分之十五，供应量为四个半月。”
场景 4 (19-25秒) - 前门和门廊的空镜素材，叠加统计卡片：24 DAYS。
“定价合理的房子大约在二十四天内售出。”
场景 5 (25-30秒) - Ava 出镜，细线规则和小型大写 CTA 卡片。
“窗口期再次倾向于卖方。考虑搬家吗？我们在秋天之前谈谈。”
```
## 编辑以达到发布完美标准
初剪是一份草稿，有两种方法可以缩小草稿与成片之间的差距。
路径一：继续对话。视频渲染完成后，聊天并不会结束。在这一轮运行中，场景 2 是个难点：背景不对。所以我们在同一个聊天中直接告诉 Agent 更改该场景的背景。它重新剪辑了场景 2，而其他每个场景都完全按批准的样子保持不变。这就是那个往返过程。如果第 17 天的“每卡评论”流程感觉对味，这运用的是同样的能力。
路径二：在 AI Studio 中打开副本。当你想亲自上手剪辑时，将副本发送到编辑器并进行完全的手动控制：脚本和文本、数字人场景、声音、时机、转场、布局、来自素材库或你上传的媒体替换、音乐，以及字幕微调。聊天处理了本次运行的修复，但当最后的百分之五（字幕挤到了脸部、剪辑慢了半拍、音乐音量）需要人手时，这就是最深层的控制所在，仅一步之遥。脚本位于左侧，场景位于画布上，而“询问 AI”、你的数字人、媒体、音乐和字幕则沿着右侧栏排列：
由此产生的工作流是：用提示词搞定 90%，用编辑搞定 10%。让 Agent 做它擅长的组装工作，然后把你的品味花在品味就是全部工作的地方。
## 最终输出
一个提示词，一个计划，用一句话改变背景。这就是最后产出的奥斯汀市场更新：
输出：
## 解锁
本系列的前十七天教会了你通过终端指导一个 Agent。Video Agent 就是把同样的工作关系中的终端去掉了：描述、审查计划、用句子指导、自己打磨最后一公里。你这一个月一直在建立的技能——足够清楚地表达你想要什么，以便 Agent 能够构建它——原来就是整个界面。
⭐ 如果这个系列对你有用，请给 HyperFrames 仓库点星。
试试看：app.heygen.com，首页上的 Video Agent 按钮
提示词指南：heygen.com/blog/video-agent-prompt-guide
第 18 天完成。还有 12 天。
## 相关链接
- [HeyGen](https://x.com/HeyGen)
- [@HeyGen](https://x.com/HeyGen)
- [3.1K](https://x.com/HeyGen/status/2080407266755510452/analytics)
- [app.heygen.com](https://app.heygen.com/)
- [star the HyperFrames repo](https://github.com/heygen-com/hyperframes)
- [app.heygen.com](https://app.heygen.com/)
- [heygen.com/blog/video-agent-prompt-guide](https://www.heygen.com/blog/video-agent-prompt-guide)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [5:38 AM · Jul 24, 2026](https://x.com/HeyGen/status/2080407266755510452)
- [3,118 Views](https://x.com/HeyGen/status/2080407266755510452/analytics)
- [View quotes](https://x.com/HeyGen/status/2080407266755510452/quotes)
---
*导出时间: 2026/7/24 09:37:21*