一句话出片：Matt Wolfe 用 Claude 自建短视频剪辑产线（含月度自我强化回路）

信源：YouTube 视频 ep4psgAFOqc（Matt Wolfe 频道，自称 “Show me your workflow, me edition” 教程/实拍向视频）。时长 1964 秒 ≈ 33 分钟；转写共 17 个 [Ns - Ms] 时间段、完整覆盖 0s→1964s，两遍通读无遗漏，本文依据英文 ASR 转写整理（附录含全量中文翻译）。讲者：频道主 Matt Wolfe，自述做视频剪辑 17 年（自己的频道+早年客户片），频道以 AI 资讯为主；判断依据为其频道公开内容与节目自述（推断级，未逐项核实）。利益披露：本期由 VidIQ 赞助，讲者自述「在对方赞助之前就长期把 VidIQ 用于这套流程」，并提供 1 美元首月推广链接（转写出现 vidiq.com/matwolf 与 vidiq.com/matthewolf 两种拼写）——赞助内容按「节目称」处理。同域旧文（背景不复述，仅互链）：消费者自建软件的宏观判断见《Matt Berman：软件成本归零，消费者自建》；同日发布的《Latent Space 上海特辑：Kill My SaaS》从供给侧论证「中等 SaaS 被克隆」，本文是需求侧的一个具体消费者的实例，两文互为印证。

先说结论

一个不写代码的 AI 资讯博主，把「短视频剪辑」这条专业流水线完整搬进了 Claude：输入一个链接，模型先用 VidIQ 扫全网同类视频的爆款规律写脚本；博主对着脚本录一段口播（A-roll）拖回对话框，一句提示词都不用加，模型加载两个自建技能（写脚本 + 剪辑），自动跑完 67 步——转写、剪掉停顿和口误、按镜头清单配 B-roll、ElevenLabs 生成配乐音效、加字幕、渲染、自检交付。产出节奏是每天 1–2 条、七 天无休，多条播放量到 137 万至 180 万级。更值得抄的是他把「每月复盘上月爆款（留存曲线、评论、转写）→ 更新技能」做成了近似强化学习的回路。他自己的定性最诚实：剪辑的边际成本坍缩必然造成内容洪水，短视频剪辑师岗位受压（拥抱者产能×5、不拥抱者出局），而创作者的价值将从「剪得比别人好」转移到「持续找到别人没见过的东西」。

一、产线全景：一个链接进去，一条成片出来

先看成品参数（均为讲者自述）：Instagram 与 YouTube Shorts 双平台日更 1–2 条；一条「完全由 AI 剪辑」的视频接近百万播放；特斯拉新车那条 13.7 万播放；Shorts 页面里 180 万、15.9 万、17.4 万播放各一条——从特斯拉那条起「每一条都是 AI 剪的」。

工作流以 Google DeepMind 的 Genie 4 视频为例：他把 Genie 4 博客链接 + 一段自己试用时录的 B-roll 素材路径交给 Claude，说明「B-roll 用官方的和我自己的」。模型随即调用 VidIQ MCP 服务在 YouTube 上检索 Genie 4 相关视频（找到 Theoretically Media 的 Tim 的视频和几条官方片），对表现最好的一批做文本分析（读它们的转写，看讲了什么），据此写出脚本，并给出预测分（这条 8/10）和剪辑注意事项。讲者明确说自己不喜欢 AI 脚本的腔调，所以只把它当「要点提纲」，几乎不照读——但脚本的价值在于它知道别人的爆款都讲了什么。

然后他开摄像机录口播（原始素材充满停顿、口误、重复——他现场放了一段自嘲样片）。把这段 A-roll 拖进对话框，零提示词，模型加载「Mr. Eflow Short Editor」技能（转写两见 Mr. Eflows / Mister Eflow，疑为其 Shorts 频道名，未核实），跑 67 步：加载 ElevenLabs 拿配乐音效、切 A-roll、准备 B-roll 与音乐床、写计划并渲染、QA 自检交付。过程中它会拉取候选 B-roll 的截图、给每段素材打时间戳以便对位、反复自查。成片里字幕、B-roll（官方素材+他的素材混剪）、配乐全部自动；连「我打字输入提示词」这种没有现成素材的镜头，它也自己用 Hyperframes 生成了一段。整条工具链：Claude（Opus 5.5）+ VidIQ MCP（爆款分析）+ ElevenLabs（音乐/音效/转写）+ Hyperframes（动效/类 After Effects）+ Blender（3D 物件）+ FFmpeg（真正执行剪辑，headless，不需要 DaVinci Resolve 或 Premiere 占用电脑）+ yt-dlp（下载网上视频做素材）。他特别注明：Codex/ChatGPT 系（转写作 Sol，疑为 GPT-6 系型号，存疑）同样能跑通这套流程。

二、从零搭起：指令、双技能与「爆款基因」提取

为了演示可复现性，他专门花 20 美元开了个全新 Claude 账号（旧账号技能与记忆齐全，「一上来就会」，不公平）。步骤：新建项目 “Shortform Video Editor”，写入角色指令——「你是世界级短视频剪辑师：把素材变成大概率表现好的脚本；清理 A-roll 的停顿与口误；从素材与全网找到 B-roll 让视频有视觉吸引力」；在 Customize 里装 VidIQ 与 Hyperframes，并连接 ElevenLabs（选完整版，才有音乐与音效）。

关键的启动提示词（本片最可复制的部分）：「用 VidIQ 分析我频道（@Mr. Eflows）的短视频，看表现最好的那些，从脚本和画面两端搞清楚它们为什么好；然后创建两个技能——一个生成同款、大概率会火的脚本，一个把视频剪成我现有视频的样子；剪辑技能可用 FFmpeg，另加 Hyperframes 和 Blender 提升观感。理想闭环：我给你一个链接→你写出大概率表现好的脚本→我录口播回传→你把它剪成我的风格：切停顿、切口误、加 B-roll、加动效、加字幕、配背景音乐和音效（用 ElevenLabs）。」外加一条专业细节：「剪辑时同时用转写文本和波形图」——只看转写容易把单词切一半，必须看波形确认声音结束后再下刀。

约 15 分钟后，模型交付两个技能（short script / short editor）和一份「你的短视频为什么能打」的分析：实体机器做出惊人动作（飞行汽车、战斗机器人、身体扫描仪）；固定钩子「This is the coolest thing I saw today」（它认出这是系列开场白）；观众会争论「真假 AI」的视频表现好（讲者只发真实内容，所以观众一旦犹豫，说明他真找到了好东西）；节奏与视觉风格：画面与 B-roll 五五开、B-roll 每 1.8–3.6 秒一切换、品红字幕。还有一条副产品洞察：VidIQ 的视觉分析把他的特斯拉、保镖片段描述成「CGI」，而脚本却声明是真实演示——「看起来像 AI 但其实是真的」的内容恰恰最能打。

三、实战一条：打分、翻车与两轮迭代

测试选题选了 NVIDIA 的 Motion Bricks（动作素材生成，画面好、适合短视频）。流程实录：贴链接→模型读页面（经 archive.org）、跑 VidIQ 比对同类、给画面吸引力打分，然后交卷：4/5 分、约 50 秒、「对不关心 AI 的人吸引力一般」、有天然争议点、关联人群关心的话题（NVIDIA、游戏、Unreal Engine），并基于他频道历史给出预估播放区间 1.5 万–3 万（「不是飞行汽车，但应该不差」）。脚本质量过关（没有「不只是 X 更是 Y」式 AI 腔）、附标题建议（「这里的每个动作都是 AI 实时生成的」「NVIDIA 刚刚取代了游戏动画师」）和下一阶段用的 B-roll 镜头清单。

录完口播拖回对话框后，撞上了三个真实坑，他都给了修法：① ElevenLabs 转写接口拒收视频、而技能不允许别的工具转写——告诉它「ElevenLabs 不行就换工具，并更新技能」；② NVIDIA 站点下载被拦——设置→Capabilities→允许所有站点；③ URL 仍被拦——开启 Create and Edit Files 与「Use the Internet while making files」，在 Allowed Websites→Manage 里手动放行被点名的 URL；跨会话权限需按模型提示开新会话重发。

第一版成片「不错但有硬伤」：没配背景音乐、人声有奇怪的压缩感、留了一处口误、节奏不够紧。他逐条反馈（含「务必转写+波形双查避免切半个词」「用 Hyperframes 加动效」「把这些改动写进技能，以后所有视频都生效」）。约 36 分钟后出第二版：配乐正常、人声干净、补了动效字幕、剪掉口误、节奏收紧——模型还顺手给自己写了两个 Python 脚本（short edit 与 GFX Kit）固化这些能力。他的总结：就是「反馈→改→更新技能」的循环，几轮之后你说一句「去把这条剪了」，它就知道全部怎么做。

四、月度强化学习回路：把留存曲线喂回技能

进阶玩法是把复盘也交给模型：对着项目说「回顾我最近十条 YouTube Shorts，找出表现异常好的离群值，基于发现给我一份改进脚本与剪辑的建议清单」。不必指定工具，它会自己调 VidIQ：拉全部视频、读频道分析、拉爆款转写、重新「看」一遍视频、读评论、看留存曲线——他把这称为「近似强化学习的回路」：每月跑一次，用上月离群值更新两个技能，下月产出自动变好。它给出的建议长这样：开场帧应是「实物真实在动」的画面（车腾空、机器人跳上车）；第 2 句要说清画面里物理上正在发生什么；资历与参数放标题卡；但书（caveat）压缩到一个从句或删掉；注意观众在第几秒流失并针对性改。这套建议随即回写进技能，供未来视频使用。

五、17 年剪辑师的矛盾自白

结尾的价值判断比工具演示更值得记录（均为讲者观点）：对「口播+信息」型短视频，AI 现在已经剪得「相当好」；长视频他仍要真人剪辑师并亲自精修。他的矛盾有三层：其一，短视频剪辑师将越来越难接活——拥抱这套的人产能约×5，不拥抱的可能失业；其二，创作门槛坍缩必然带来内容洪水（「任何人都能不停地量产视频」），他不喜欢，但「说出来几乎像在搞行业守门」——他靠内容吃了 17 年饭，自认没资格；其三，他清楚自己的价值已经转移：如今大量时间花在找信息、读文章、发现值得分享的东西上——「我在短视频上的竞争力不是剪得比别人好，而是持续把别人可能没见过的好东西放到他们面前」。所以尽管「讨厌自己正在成为内容商品化的推手之一」，他还是决定把方法全盘公开：这件事无论如何都会发生，不如把知识分享出来，倒逼自己往上走一层。

机制、二阶影响与读者启发

1. 技能沉淀 = 个人产线的复利资产。变化：一次性的工作流被固化为两个可迭代技能（script/editor）+ 月度数据回路（VidIQ 留存曲线/评论/转写当奖励信号）。机制：把「品味」写成 spec，让模型每次执行时都加载、每次反馈后都更新——个人创作从「每次手工」变成「一句话触发」。二阶：月度回路使技能与频道数据共同进化，后来者即使拿到同一套提示词，也拿不到你的历史数据。启发：任何重复性专业工作都可以拆成「分析→脚本→执行→复盘→更新技能」五段；奖励信号用自家历史表现，比通用榜单更贴身。
2. 剪辑边际成本坍缩 → 内容洪水 → 策展升值。变化：一条 50 秒短视频的剪辑成本从「人工数小时」降到「拖入文件+36 分钟两轮反馈」（节目称）。机制：FFmpeg headless 剪辑 + 素材自动检索（yt-dlp/VidIQ）+ 生成配乐，把专业软件栈整条搬进对话。二阶：产能过剩使分发位与「可信的发现者」更值钱（讲者自己的定位转移即是证词）；拥抱者×5、拒绝者出局。观察指标：短视频剪辑师接单价、平台内容供给增速、头部创作者的更新频率。启发：与其抵制商品化的环节，不如把自己移到「选题与策展」这层——讲者的原话是「倒逼自己往上走一层」。
3. 消费者自建的又一个实锤（需求侧）。变化：非程序员靠 Claude+MVC 生态（VidIQ/ElevenLabs/Hyperframes）+ 开源工具（FFmpeg/yt-dlp/Blender）自建了一条原本需要 DaVinci/Premiere 许可证与外包剪辑师的专业管线。机制：每样「中间层专业软件」都被开源等价物+agent 编排替代——与同日《Kill My SaaS》一文「模型点开每个界面就能复刻」的供给侧判断互为印证。启发：判断一个 SaaS 的暴露度，看它的功能能否被「开源等价物 + 一段编排」覆盖，以及它的客户里有多少像这样会自己动手的人。

未采用条目及原因

• 「My name is Skrillex」音乐样例、二维码摆放玩笑、结尾感谢语：纯花絮，无信息量，未入正文（翻译附录保留）。
• vidiq.com/matwolf / matthewolf 两种拼写、@Mr. Eflows 频道名：无法核实哪个为准，入校正表存疑，正文已注明。
• 第二账号 20 美元/月、演示后即取消：仅操作细节，正文一笔带过（「花 20 美元开新号」）。

转写校正表

转写原文（修正版）

[0s - 120s] So this is wild. AI is finally getting really good at something that I didn’t think it would ever get that good at. At least not as quickly as it did. AI can finally edit my videos for me. I give it some A-roll of a talking head of me. I give it the prompt, go edit this for me, and it comes back with a video that I’m pretty happy with. Now, right now, I’m pretty much just using it to edit short form videos, but I’ve become absolutely obsessed with using AI to edit this style of video. If you check out my Instagram page, you’ll notice I’ve been posting at least one new short a day, sometimes even two a day, seven days a week. And some of these videos have been doing quite well. Like this one’s almost hit a million views. This one was entirely edited with AI. Here’s another one about Tesla’s latest car that’s done 137,000 views. And if you look at my YouTube Shorts page, very very similar story. I’m posting all the same videos to Instagram and YouTube Shorts. This one down here has done 1.8 million views. Here’s another one that did 159,000. Another one that did 174,000. Every video that you can see from about this one up, this Tesla one here, I used AI to edit these videos. So in this video, I want to walk you through what AI is actually capable of when it comes to editing, and then break down a step by step workflow that you can follow to set up your own editing process using AI. This is show me your workflow, me edition. So for my workflow, I’ve actually been using Claude with the Opus 5.5 model, but this absolutely works using Codex and the ChatGPT models. Like Sol works great for this as well. So here’s a quick breakdown of what the workflow I built does. So for this example, I was making a video about Genie 4. I gave it a link to the Genie 4 blog post so it can grab the details from there. I also gave it a little extra context. There’s a lot of videos on.

[120s - 240s] That page. I also recorded my own B-roll of me testing it. I gave it a link to where it can find that on my computer, and then I said, “Let’s use B-roll from both the original source and my B-roll.” So now it has access to look at this website, where it can grab all the details it needs to know about Genie Four. There’s a bunch of videos on this web page that it can use for B-roll inside of the final video, and I linked it to a video on my hard drive where it can show some of the B-roll that I generated when I had a chance to play around with this tool. And then what it did was it fired up the VidIQ MCP server. I’m going to talk about them a little bit more during the setup process, but it looked on YouTube for other videos talking about Genie Four. It found Tim from Theoretically Media’s video about Genie Four. It found a couple official videos about Genie Four, and it basically found you know the best performing videos that were out there on Genie Four to help inform like what content should go into the video itself. So once it did its analysis using VidIQ, you can see here that it actually did a text analysis. So it looked at the best performing videos, looked at the transcripts of those videos to find out what they were talking about, and then used that to inform this little script portion. It then writes me a script. Personally, I don’t totally love AI generated scripts. I feel like you could still tell they’re AI generated, so I use the script as sort of like a guideline. I almost never read it word for word, but it did give me a script here. And the nice thing about the script that it gave me is that it’s going to inform me of all the talking points that I should try to inject in that video because it knows what did well from other people’s videos. It also gives me a score of how it thinks it’ll perform. So it thinks this one’s an eight out of ten. It gives it some notes here about the actual editing. And again, all that I started with to get to this point was a link to Deep Mind, and then sharing a link to where some of my own existing B-roll was. If I was making a video about Genie Four, there’s a ton of video already that Google’s put out, so I don’t even need to make my own B-roll. But because I did have early access, I opted to use some of it here.

[240s - 360s] In most of my videos that I do, if I don’t have access to the thing, I don’t give it any extra B-roll. I just give it a link to the resource where it can learn about it, and then it goes and does more research using VidIQ here. So once it’s got the script, then I’m ready to record my A-roll. That’s when I flip on my camera and just speak into the camera about the thing I’m talking about. Again, I use this script as a rough guideline, never exactly word for word from the script. So here’s what the raw video looks like when I gave it to Claude. You’ll notice there’s still pauses. I make mistakes and repeat myself. Now, I don’t really think this. Now, I don’t really think this looks like Rainbow Road from Mario Kart that much. As you can see, it needs edits. It’s got long pauses. It’s got me repeating myself and making mistakes. But all I have to do is here’s the A-roll I generated and a link to that video, and it knows what to do. You can see it ran, you know, 67 steps here. If I open up all the steps it ran, it loaded my skill, the coolest thing. Edit, set up the cut and A-roll, prepare B-roll clips and music bed, write a plan and render, QA and deliver. And you can see it went through all of these tasks down here. It loaded up ElevenLabs. That’s where it got the background music and any sound effects that might be in it. It set up and cut the A-roll. It then it moved on to preparing B-roll clips and the music bed, and then it did the thing with ElevenLabs. Then it loaded Claude in Chrome, tried to run a computer command and failed, but then you know clearly figured out workarounds, pulled some of these images, and what these images are are actually screenshots from different videos that it was looking at for B-roll. So it knows exactly what to pull in for B-roll. This is actually from the B-roll I gave it here. You can see that it’s got timestamps of every little scene inside of this B-roll, so it knows how to line it up. And you’ll see these screenshots throughout. It just keeps on double checking its work over and over again. Down here we can see where it generated music for underneath the video.

[360s - 480s] My name is Skrillex. That all used Eleven Labs, and then it gave me a summary of what it built, and we have the final video down here. When I press play on the video, this is what the video looked like. This is the coolest thing I saw today. I typed a giant gorilla enters the city, and then it just walked right into the street I was standing in. This is Genie 4 from Google DeepMind. Now, if I scrub through it real quick, some things to notice: these subtitles here, it added those. I didn’t do anything to add those. It added all of the B-roll you see. This is actually B-roll that I gave it, but then some of the B-roll, when I get a little bit further into it, like the B-roll of the giant tentacle monsters while these people are on the scooter, that was actually B-roll that came straight from the DeepMind website. The background music you heard was generated by Eleven Labs. This video here of it typing in an environment and a character, it actually used Hyperframes and generated that itself because I didn’t have good B-roll of me typing that. And this video is really just scratching the surface. I’ve improved the workflow even more since this one. You can actually have it go and generate 3D objects in Blender and pull those 3D objects in. You can have it generate charts and graphs as you’re talking about things and add little animations over the B-roll. We’re gonna get into all of that, but this is just a quick teaser of what you can do with it. Okay, now to really properly show this off, I feel like I need to create a brand new Claude account. If I was to go use my existing Claude account, it has all of my skills already in it. It’s got all of my past memory. It’s gonna sort of jump right to being good. So for this tutorial, I’m gonna bite the bullet and purchase a second Claude account and walk you through how to do it on a from scratch Claude account. And I’m just gonna use the monthly plan at twenty dollars a month. I don’t want to do yearly because after I do this demo, I’m gonna cancel this version of Claude. All right, and as you can see down here in the bottom, it’s a pro account, but you can also see there is zero history. This is a brand new, fresh, blank slate Claude account. So step one, we’re gonna create a new project. We’re gonna call this project Shortform Video.

[480s - 600s] Video editor. Now up here for the instructions, I’m going to go ahead and click Add, and this is just a sort of general overview of what this project is for. So I’ll put something like, “You are world-class short-form video editor. You take the source material and turn it into a script that is likely to perform well. You clean up A-roll video to cut out pauses and mistakes. You edit in B-roll from the source material and from what you find around the web to make the video highly visually appealing. I’m also going to tell it you can use all the tools at your disposal, like VidIQ, Blender, Hyperframes, FFmpeg, yt-dlp, and as I go through the video, I’ll talk about all of these. But basically, VidIQ is going to look through existing videos either on your channel or other people’s channels to find what has a high likelihood of going viral, and it’s going to analyze that. Blender will help you make sort of 3D objects and even some animations that can go into your video. Hyperframes does almost like After Effects-style, real basic editing. FFmpeg is actually going to do the editing, so you don’t even need to use something like DaVinci Resolve or Adobe Premiere. We’re doing this all headless, so it’s not actually going to go use another editor and take control of your computer. It’s just going to do the editing behind the scenes, and then yt-dlp is actually a script that allows it to download videos from the web. So when it’s using that, it can actually download YouTube videos and Instagram videos and videos that it finds and use that for B-roll. So I’m going to go ahead and save the instructions here, and then over under Customize on the left, we want to make sure we install VidIQ. This is sort of like the secret sauce to everything I’m doing here. So I did a quick search, and VidIQ is right here. So I’ll go ahead and click on that, and we’ll connect this to Claude. So we get this dialog here. I’ll click Continue Connecting. It’s going to bring me over to VidIQ, and I will go ahead and authorize my VidIQ account. Now VidIQ has graciously sponsored this video, but I want to tell you I’ve been using VidIQ for this process long before they actually agreed to sponsor this video. Like I’ve been using VidIQ in my video creation process. I told VidIQ, “Hey, I’m going to make a video about how I edit.”

[600s - 720s] My videos and how VidIQ is a big part of it, and they said, “Hey, let us sponsor that video.” So that’s exactly how that came to be. So VidIQ is a big part of my workflow, and you’ll see why as I get deeper into this video. But also, when I told them I’m making this video, they said, “Hey, let us set you up with a page where you can offer it for a buck for their first month.” So if you head over to vidiq.com/matwolf, you can actually use the VidIQ MCP, which you’ll see throughout this video for just one dollar. If you’re watching on TV or something, they also put this QR code that you can use. Thank you so much to VidIQ for sponsoring this video. I think it’s going to make a lot of sense as you watch the rest of this video as to why I like VidIQ so much. Now, while we’re still in this Customize tab over here, let’s also add in Hyperframes. This one’s going to help with some of the animations in the videos. So we’ll go ahead and select this, and we’ll connect this one to Claude as well. It’s going to send you over to your browser once again to confirm, and once you do, it will be connected. All right, let’s jump back into our shortform video editor here, and there’s a couple ways that we can get started. If you already have an existing YouTube or Instagram account, you can have VidIQ go and look at that and analyze which videos performed well, and have this tool make more videos like that. That’s exactly what I did. If you don’t have one, you can tell it some examples of YouTube channels that you really like and might want to take some inspiration from. Since I already have a YouTube channel, I can say, “Use VidIQ to analyze my shortform videos on my channel, and I’ll just kind of put in parentheses at Mr. Eflows. Look at the best performing shorts and figure out what makes them perform well. Look at both the script and the visuals, and then create a skill that will generate similar scripts that are likely to go viral, and another skill that can edit videos to look like the existing videos. The editing skill can use FFmpeg to edit, as well as Hyperframes and Blender to add some additional visual appeal and polish. Ideally, the end result is that I give you a URL to something interesting that I want to make a short about, then you write a script that is likely to perform well. I then record my talking head.

[720s - 840s] A-roll of that script, feed it back to you, and then you edit that video to look like my existing videos. Cut out the pauses, cut out the mistakes, add B-roll, add animations, add subtitles, add background music and sound effects. We can use ElevenLabs or background music and sound effects. And one thing that I do like to do when telling it to edit is this: when editing and doing initial cuts, use both the transcripts and the waveforms for editing. And the reason I tell it that is sometimes it’ll read the transcripts, think it needs to cut it off, but it’ll accidentally cut you off like mid-word. So it needs to also look at the waveform to make sure that it’s cutting after the sound ends and not when it just thinks the word is. I’m going to go ahead and submit this, and it’s going to start building those skills. And it’s going to work for a little while to actually create these skills. But what we’ll see it do first is it’s going to load VidIQ here and do a bunch of analysis on my existing videos. To speed this up, I’m going to come down here to where it says manual and also set this on auto mode, so it just keeps on running. And I’m going to click always allow, so it just keeps on doing what it needs to do. And we can see down here it’s opened VidIQ and it’s looking at my videos right now. You can see it found the outliers. This one did 3 million views, 1.9, 1 million. So it’s analyzing what made these videos work so well for me on my channel. Now, as it does this process, I will say this is going to be a very iterative process. It’s going to come back. It’s going to give us both skills: the scripting skill and the video editing skill. The scripting skill will probably be pretty dialed in from like the first go. The editing skill will probably give it our A-roll. It will probably mess it up. It’s probably going to edit it in a way that we’re not totally happy with. But then we’re going to tell it what to fix, and then once we tell it what to fix, we will then tell it to then update the skill, and we’ll keep going back and forth until the edits get better and better and better. And eventually, you get to the point where you just say, “Go edit this for me,” and it knows exactly what to do. Okay, so that worked for about 15 minutes, and we can see.

[840s - 960s] That it looked at my top-performing videos and it created two skills, so the short script and the short editor. I’m going to go ahead and save both these skills here, and then we can review what it did. So it analyzed the shorts and drafted both skills. So we can see here what makes your shorts perform. This is what it figured out by looking at all the VidIQ data. So physical machines doing something surprising, things like flying cars, combat robots, body scanners. The hook that seems to work is this is the coolest thing I saw today, which is the name of my series. So it knows that that’s what I start every series with. Videos where people actually debate whether or not the thing they’re seeing is real or AI. I only post real stuff unless I say otherwise. So it’s kind of funny, but if people have to wonder, that means I probably found a really cool piece of tech. It also looked at my pacing and the visual style, 50/50 split with B-roll on top, magenta captions, B-roll changing every 1.8 to 3.6 seconds. Now it does still need a few things from me to get this whole workflow working. It does tell me what I need from you: connect a folder. So I need to add a folder for it to use to put the video files in and things. I need to add my API key for Eleven Labs, and then it wants me to do a test recording. We can see here some more analysis. VidIQ’s visual analysis described the Tesla and bodyguard footage as CGI, while the scripts present them as real demos. So again, videos where people go, “I feel like that’s AI, " but it’s actually real footage seems to do pretty well. So the first thing it wants me to do is connect a folder. So I’ll click this little plus down here, click Add Folder. I create a folder called Claude Editor. So I’ll go ahead and open that as our folder, and I’ll just go ahead and send that so it knows the folder we’re working with. I’m going to select Don’t ask again for this folder on this device. We’ll allow that. Now it wants me to allow network access for Eleven Labs. So for that, I’m going to come to Customize, and then I’m just going to search for Eleven Labs. I’ll select this and connect Claude. It’s going to give me a page that looks like this, asking me to connect in my browser. I’ll click Authorize.

[960s - 1080s] But then it’s going to ask me if I want to install this version or the full version, and I want to install the full version because that one’s going to get me access to music and sound effects and things like that. We’ll go ahead and click continue on this screen here, and then it’s going to have us sign into our Eleven Labs account. So I’ll click add. We’ll connect again, and we’ll go ahead and authorize everything. And now we are connected. So we’ve got our folder added. We’ve got Eleven Labs connected, and now it wants us to send a test recording. Now it does say one raw A-roll take with its script will show how well the cuts work on your real voice, but I’m not going to give it a script because I want it to pull the script itself if it doesn’t have one. So I’ll tell it when I give you a video use Eleven Labs to get the time stamped script, and it should update the skill to let it know to grab it from Eleven Labs. So we can see it went ahead and updated the skill. So I’ll go ahead and click update, update, update. I don’t know why I have to click three times, but now our skill is updated, knowing that if I just give it a raw video with no transcript, it’s going to use Eleven Labs to grab that transcript. I’m going to click back on our project here, our main project folder, and under context, I’m going to go ahead and add the folder that we made earlier, just so it always knows it can use that folder. So this one I called Claude Editor. I’m going to go ahead and open that and always allow, so it knows it can always use that folder as well. And now theoretically, we’re set up where I could give it a link to anything, and it will go and create a script for me. Then I can record the A-roll for that script, give it back the A-roll, and then it will edit the whole thing for us. Again, it probably won’t be perfect the first time around. Well, nobody’s perfect. But it’ll get pretty dang close the first time, and then we just give it feedback. So let’s find something real quick. So I recently came across this motion bricks out of Nvidia, and I thought this one would be a cool one to make a short out of because it’s got a lot of really cool visuals that we can pull from. So it seems like it would really grab attention as a short. So theoretically, all I’m going to need to do: copy this URL here, open up Claude, I’ll go to our short form editor project here, but create a new window, and I’m just going to paste.

[1080s - 1200s] This URL in here. That’s it. I’m gonna hit submit. We can see it’s reading the pages. It just loaded our skill, Mister Eflow Short Script. So it’s using that short script skill. It’s reading the archive.org page, and now it’s loading up VidIQ again. It’s reviewing videos that were similar on VidIQ. It’s scoring the animation’s visual appeal. So it’s actually looking through what’s on the page to figure out what the most visually appealing aspects are to use as B-roll in the next phase. And here’s what it came back with. It scored it a four out of five. It’s about fifty seconds. You can see it in a glance. It passes that someone not into AI would stop. So it doesn’t think it’s strong if you’re not into AI. But good footage exists. There’s natural debate around it. Links to something people care about. Nvidia, video games, Unreal Engine, etc. It gives me an estimated view range based on what it found on VidIQ and everything it has access to. So game and 3D world shorts on your channel land around 15 to 30,000 views. I’d expect this one to fall about that. It’s not a flying car, but it thinks it should do decent. And then it gave me the script. Now it’s not following my other script formula that starts with “This is the coolest thing I saw today” because the skill hasn’t been taught yet to write the scripts in that format. I’m not gonna do that for this one because this is a brand new script designer. My other skill that I’ve already built already writes scripts exactly the way I like them. So what I would do now is I would read this script, and if I like the way this script sounds, I would just say, “Cool, I’ll move on and record this script as A-roll.” If I think the script needs work, I’m gonna give it some feedback. Quite honestly, I just read this script, and it’s pretty solid as is. A lot of times you get some like AI isms in it, like it’s not just this, it’s that, and things like that. But this one doesn’t have any of that. It doesn’t read like a super AI script. So it did a good job of analyzing my previous scripts and making one that sounds fairly similar. It gave us some title suggestions. Every move here is AI generated. AI is animating this game live. Nvidia just replaced game animators. And then it also made a shot list for B-roll so that when I move on to

[1200s - 1320s] The next phase, it knows what B-roll to grab and where to put it. So my next step would be to go and record this as A-roll, me talking into a camera, and then just feed it back. So I’m going to go ahead and do that real quick. Okay, so I went away and recorded the A-roll of me basically reading this script here, almost word for word, but I ad libbed a little bit. Here’s what the raw footage looks like. Look at it, walk, jog. Look at it, walk, jog, and then run. Look at it, walk, jog, and then break into a full run. So as you can see, pauses, mistakes, do overs, it’s all in there. And here’s where the next step is really cool because all I have to do is take this video that I just recorded, this video file, drag and drop it in here. I don’t even need to give any other prompt. I’m just going to submit that video, just like that. We can see it loaded our skill, the Mister Eflow Short Editor. It’s transcribing our rough cut, getting B-roll and visuals, music, render, and review. So it’s going to do all of this for us now. I just step away and wait. Okay, so I actually ran into a little issue here, and it stopped before it finished the edit because of a few blocks. So if you run into this, this is how you fix that. It says that Eleven Labs is not accepting my video to transcribe, and apparently the skill won’t let something else transcribe it. So we need to fix that. So I’m going to go ahead and tell it if Eleven Labs isn’t working for transcription, we should use an alternate tool. Please update the skill. It’s also saying downloads from Nvidia’s site are blocked on the computer. Now the fact that it won’t read Nvidia’s website, that should be fixed by me going to settings, capabilities, and then selecting allow all sites. So now it should be able to read all sites that I tell it to. So Nvidia site should work now. Okay, so I went ahead and updated the skill. So we’ll go ahead and make sure those updates are added in so it doesn’t stop if Eleven Labs isn’t working, but it’s

[1320s - 1440s] Telling me I need to use a new session in order to access the sites, so it’s telling me what to prompt a new session with. Edit this into the Motion Bricks short. Then it already shows the URL of where the script is, so we’ll go ahead and copy all of that. We’ll just start a new session, and then I’ll drag the video file into this new chat, and hopefully all of our permissions are going to work now. Okay, so it actually gave me one last little issue here. It was saying these URLs were blocked, and I went, I thought I unblocked all this stuff, but I haven’t figured out now. So if you go into your settings, so you click on Capabilities, and then down here where it says Create and Edit Files, turn that on. Turn on the thing that says Use the Internet while making files. Make sure that is on, and then under Allowed Websites, click on Manage here, and this is where you plug in those URLs that it’s saying are blocked. So if you see URLs are blocked, go to that screen I just showed you and unblock them. So I did that, and I literally prompted, isn’t working now, and then it went off and made my video. So I haven’t watched this yet, but this is what it did on that first attempt. Every move this character is making is being generated by AI live. This is Motion Bricks from NVIDIA. Look at it walk, jog, and then break into a full run. It vaults over a bench, sits down, picks up a sword, and there’s no handmade animation stitching all that together. So it’s not bad, but there’s definitely some issues with it. It didn’t do the background music. The audio sounded a little bit weird, like there was some extra compression or something on my voice audio. I noticed one of the mistakes that it didn’t cut out, and I’d like it to tighten up the gaps in my speaking even more. So I’m just going to give it a bunch of feedback. So I gave it the feedback. Let’s get the background music working. My vocal audio sounds weird with some weird EQ or compression. Let’s tighten up the gaps with me speaking. It also left in some mistakes where I misspoke. Fix that. I gave it additional instructions to be sure you both transcripts and the waveforms to ensure you don’t cut me off mid-word. Unless use Hyperframes to add some animations. Didn’t add any extra like animation stuff in there. And then I finally told it make sure these changes are updated in the skill so they’re applied to all future videos.

[1440s - 1560s] Videos as well, so let’s have it do another pass with this feedback. Okay, so it’s been about 36 minutes, and we have a new revision. It’s actually still working, but right now it’s just updating the skill files so that it knows how to do this stuff in the future. We can see it even created some Python scripts here so that it can better accomplish these goals in the future as well. It created a short edit Python script and a GFX Kit Python script. Let’s go ahead and take a look at this new version here. Every move this character is making is being generated by AI live. This is Motion Bricks from NVIDIA. Look at it walk, jog, and then break into a full run. It vaults over a bench, sits down, picks up a sword, and there’s no handmade animation stitching all that together. Usually, game animators need to build every single one of those transitions by hand. So that was a much better version. The audio sounded better. We had some background music. It added some more graphics and animations overlaid, and it you know tightened up the audio gaps and removed the mistakes that I heard the first time around. So that is really how you do the video editing. You keep on going back and forth and back and forth, and we can see it updated our skill based on this new feedback. So I can go ahead and tell it to yes, go ahead and make sure that’s in the skill from here on out. But you can even take this further. You can add in Blender and have it add 3D objects. In fact, let me jump over to my other account and I’ll share some additional tips. In this video that I made for this Hynomas world model here, if I preview this video here and fast forward a little bit, you can actually see an instance where it jumped into Blender, created an animation of a GPU, and then used Hyperframes to add some extra text over it. So you can tell it to jump into Blender and make 3D objects if you don’t have the B-roll. So if you’re making a video that has a really cool concept, but there’s not great B-roll around the web for it, tell it to use Blender and generate graphics for you and animate them, and it will actually do that as part of the B-roll. Here’s something I recently came across from Monique Price, who I believe edits for NetworkChuck here, but she talks about how.

[1560s - 1680s] She gave Claude a sequence that she had in mind, and it created this. And when we watch this video, you can see it probably made that toilet in Blender, and then the actual text and animations on top. It probably used something like Hyperframes, or even possibly did that in Blender. So this shows you the type of quality of animations you can start to put into your videos here, just by prompting Claude. Now, don’t get me wrong; there will be a lot of back and forth, and a lot of iteration to get it exactly where you want it, but it could do probably most of what you can think of in the video. Now, another thing I like to do is inside of my short form video creator project here is I’ll tell it to review my last ten YouTube shorts for outlier videos that performed well and give me a list of suggestions based on your findings for how I can improve my scripting and editing skills. Now, I won’t even need to tell it what tools to use, and by default, it’s likely going to call upon that VidIQ MCP. We can see it already pulled in VidIQ here. I didn’t even need to tell it to do that. You just have the VidIQ MCP installed, and it will review all of your short form videos. If I click this little arrow here, I can see this sort of step by step of what it’s doing. So it used VidIQ to pull all the videos from my channel here. It’s looking at my channel analytics. It’s pulling the transcripts from the videos that it pulled that it found to be the best performers. It’s actually watching the short form again. I know I told you this was sponsored by VidIQ, but like it really is one of the most powerful tools if you’re using Claude for editing videos. It’s also reading comments on the videos, so if people are giving feedback on your videos, it’s going to pull that in as well. It can look at the retention curves on videos, and the idea behind this exercise is that you’re almost creating this like reinforcement learning loop to make your videos better and better every single time. So you run this process maybe once a month or so. Tell it to look at your last month’s worth of videos, find the outliers that performed well, use that information to update your skills, and.

[1680s - 1800s] And then all of your future videos will then be a little bit better based on these findings. And if you keep on doing this every month, your videos are only going to get better and better. The scripting is going to get better, the editing is going to get better, and the skills are going to keep on learning how to make themselves improve because you’re doing this loop of improving the video process over and over and over again. And we can see here’s the feedback that it gave. It’s looking at things like the opening frame opens on a car in midair, opens on a robot jumping onto a car, opens on a close up of a guy’s face. Genie 4 cuts to the prompt typing screen at two seconds. It tells us where people are dropping off, so we can improve certain areas. So here’s the suggestions for scripting. Line two says what’s physically happening on the screen and playing words in the middle. Put something like here it is doing X. Put credentials and specs on title cards. Keep caveats to one clause or cut them. Editing suggestions down here. The first frame should always be real footage of the thing, like in action. It also gives some suggestions for title names. So as you can see, we’ve got a lot of feedback here that I could feed back to my skill, and now future videos are going to use this feedback and ideally get better and better results. Now my final thoughts on this whole process are: I have very mixed feelings about this. I’ve been a video editor myself for seventeen years now, both editing my own videos and also editing clients back in the day. But when it comes to editing this style of video, like talking head informational short form video, I feel like AI does a pretty good job at it now. Now when it comes to long form, I still like the idea of having an editor and doing some of the edits myself and like really getting in there and fine tuning the way I like it, the old fashioned way. But when it comes to this kind of short, I feel like it’s pretty easy to just knock them out quickly, letting AI do a big chunk of the editing work for you. And the reason I have mixed feelings about that is I do think you know short form editors are going to have a harder and harder time.

[1800s - 1920s] Getting work, the ones that embrace this might be able to do like you know five times as much work, which is great for those that embrace it. But there’s also going to be a lot of editors that could be out of work. But I think the bigger implication of this is that we’re going to see a lot, a lot, a lot of flooding of content. It’s gotten so easy, and the barrier to entry to create, especially short form content, has gotten so low that like anybody’s going to be able to just crank out videos constantly. And I don’t really like that. But as I say that, it almost kind of feels like gatekeeping a little bit because you know I’ve made a career out of content for the last 17 years, so I don’t know. Like I just don’t know how to feel about all of this. I spend a lot of my time these days finding cool information, reading articles, learning about cool tech, finding things to share. And I feel like that’s where more and more of my value has come in, especially when it comes to this short form content. Isn’t like trying to out edit people on the shorts. It’s trying to be somebody that’s constantly finding cool stuff and putting it in front of people that they may have not otherwise seen. But I wanted to put this all together, put it out there. I feel like if people take what I’m doing here and start using it, and we see more and more of this type of video, it’s only going to cause me to sort of like up level what I do and find new ways to stand out because this kind of thing is becoming commoditized. So might as well put what I know out about it, show you how it’s done. I hate the fact that I’m kind of becoming part of the cause of the commoditization, but at the same time, it’s going to happen regardless. So might as well share my knowledge. I want to thank VidIQ once again for sponsoring this video. It kind of came very, very naturally for them to be a sponsor. I told them I was making this video, and they said, “Hey, let us sponsor it.” And I said, “Yeah, let’s do that,” because I was going to talk about VidIQ whether they sponsored it or not. But again, thank you so much to them for sponsoring it. Head over to vidiq.com/matthewolf to actually use VidIQ for just one dollar, or you can again scan the QR code that should be on the screen.

[1920s - 1964s] Like right here, maybe, or I don’t know. Editor should put it somewhere there, or they’ll put it somewhere else to like mess with me. I don’t know what what there should be a QR code on the screen. Thank you so much for hanging out, nerding out with me. This was probably a bit longer than I thought it would be, but I wanted to make kind of the exhaustive guide, and also any of the little roadblocks that I ran into, I wanted to point them out. So if you run into them, hopefully you can get over them as well. And also as a quick reminder, you can do this with OpenAI inside of Codex as well, and I have done it with that. But I’ve made a lot of videos about OpenAI lately, so for this one, I wanted to jump in and play around with Claude to get it to work. But again, you could use either one, so whatever plan you already have, you don’t need to go get a different plan. Thank you for hanging out, nerding out. I really, really appreciate you, and hopefully I’ll see you in the next one. Bye bye.

中文翻译（全量）

[0s - 120s] 这太野了。AI 终于在一件事上做得非常好——我原以为它不会这么快做到这么好：AI 终于能替我剪视频了。我给它一段对着镜头说话的口播（A-roll），给它一句指令「去帮我把这个剪了」，它回给我的视频让我相当满意。现在我只用这套方法剪短视频，但我已经完全迷上了用 AI 剪这类视频。去看我的 Instagram，你会发现我每天至少发一条、有时两条 short，一周七天不带停。其中一些表现相当好：这条快到一百万播放了，完全由 AI 剪辑；另一条讲特斯拉最新车型的，13.7 万播放。再看我的 YouTube Shorts 页面，情况几乎一样——我把同样的视频同步发到 Instagram 和 YouTube Shorts。下面这条 180 万播放，另一条 15.9 万，再一条 17.4 万。从大约这条往上、包括这条特斯拉，你看到的每一条视频都是我用 AI 剪的。这期视频我会带你走一遍：AI 在剪辑上到底能做到什么，然后拆解一套你可以照做的分步工作流，用它搭起你自己的剪辑流程。

[120s - 240s] 这就是「给我看你的工作流——我自己版」。我的工作流用的是 Claude 的 Opus 5.5 模型，但用 Codex 和 ChatGPT 系模型也完全行得通（转写提及 Sol，存疑）。快速过一遍我搭的这套工作流能干什么。以做 Genie 4 的视频为例：我给它 Genie 4 博客文章的链接，让它从那里抓细节；我又给了点额外背景——那个页面上视频很多；我还自己试用时录了一段 B-roll，把电脑上的素材路径给它，并说「B-roll 用原始来源的和我自己的」。这样它既能访问那个网站拿 Genie 4 的全部细节（网页上有一堆视频可以当最终成片的 B-roll），又能链接到我硬盘里那段我试用时录的素材。接着它启动了 VidIQ MCP 服务——安装环节我会再多讲几句——它在 YouTube 上检索其他讲 Genie 4 的视频，找到了 Theoretically Media 的 Tim 的视频、几条官方视频， basically 把市面上表现最好的 Genie 4 视频都找了出来，用来决定这条视频该装什么内容。用 VidIQ 做完分析后，它做了一次文本分析：看表现最好的那批视频、读它们的转写、搞清楚它们都在讲什么，再据此写脚本。它确实给我写了脚本。说实话我不太喜欢 AI 生成的脚本——你还是能听出 AI 味——所以我把脚本当「指导提纲」用，几乎从不照读。

[240s - 360s] 但这份脚本的价值在于：它知道别人的爆款视频里哪些点讲得好，会提示我应该往这条视频里塞哪些 talking points；它还给了一个预测分（它觉得这条 8/10）、一些剪辑注意事项。而达到这一步，我给它的全部东西就是一个 DeepMind 的链接，加上我已有素材的路径。多数情况下，如果我没拿到产品的早期试用，我就不给任何自己的 B-roll，只给它资源链接让它自己去学，然后用 VidIQ 做更多调研。脚本到手，我就可以录 A-roll 了：开摄像机，对着镜头讲。照例，脚本只是粗提纲，我从不逐字念。这是我给 Claude 之前的原始素材样子：有停顿、有口误、有重复——「现在我真的不觉得这看起来多像马里奥赛车的彩虹之路」——可以看到它需要剪：长停顿、重复、错误都在。而我要做的只是：这是我的 A-roll、这是那段视频的链接，它就知道该干什么了。它跑了 67 步。展开步骤列表：加载我的技能（the coolest thing）、剪辑、切割 A-roll、准备 B-roll 片段与音乐床、写计划并渲染、QA 与交付，一路做下来。它加载了 ElevenLabs——背景音乐和可能用到的音效都从那里来；它切割好 A-roll，接着准备 B-roll 片段和音乐床，完成 ElevenLabs 那步；然后它加载 Claude in Chrome，试了一个电脑命令、失败了，但显然找到了绕行办法，拉了一些图片——这些图片其实是它为选 B-roll 而看的各种视频的截图，所以它确切知道该拉哪些素材进来。这是我给它的 B-roll 里的：你能看到它给这段素材里的每个小场景都标了时间戳，所以知道怎么对位。这些截图全程都在出现，它一遍又一遍地自查。下面这里是它为视频生成的配乐。

[360s - 480s] 「我的名字叫 Skrillex」（音乐样例）——那全部来自 ElevenLabs。然后它给了我一份它构建内容的总结，最终视频就在下面。按下播放，成品长这样：「这是今天我看到最酷的东西。我输入『一只巨型猩猩进入城市』，它就径直走进了我站的这条街。这是 Google DeepMind 的 Genie 4。」快速拖一下进度条，注意几件事：这些字幕是它加的，我一个字没动；你看到的所有 B-roll 都是我给它的素材；再往后一点，比如人们骑摩托车时出现巨型触手怪的那段 B-roll，是直接从 DeepMind 官网来的；背景音乐是 ElevenLabs 生成的；那段「在环境里输入提示词和角色」的视频，是它用 Hyperframes 自己生成的，因为我没有自己打字的好素材。而这只触及皮毛——这条之后我又改进了工作流：你可以让它去 Blender 里生成 3D 物体再拉进视频；可以让它在你讲到相关内容时生成图表、在 B-roll 上加小动效。这些我们都会讲到，这只是个预告。好，为了好好展示这套东西，我觉得我需要一个全新的 Claude 账号。如果用我现有的账号，它里面已经装了我全部技能和过往记忆，会「一上来就直接很好」。所以这次教程我咬咬牙买了第二个 Claude 账号，带你从零账号做一遍。就用每月 20 美元的月付——我不想年付，因为这个 demo 做完我就把这个号取消掉。可以看到底部显示这是个 Pro 账号，但历史记录为零：一张全新的白纸。第一步，创建新项目，命名为 Shortform Video Editor。

[480s - 600s] 上面的 Instructions 里点 Add，写一段项目总述，比如：「你是世界级短视频剪辑师。你把源素材变成大概率表现好的脚本；你清理 A-roll 视频、剪掉停顿与错误；你从源素材和全网找来的素材里剪入 B-roll，让视频有极强的视觉吸引力。」我还会告诉它：你可以使用手头所有工具——VidIQ、Blender、Hyperframes、FFmpeg、yt-dlp，后面我会逐个讲。basically：VidIQ 负责检索你自己或别人频道上已有的视频，找出高概率爆款并做分析；Blender 帮你做 3D 物体甚至一些可入片的动画；Hyperframes 做近似 After Effects 那种很基础的动效；FFmpeg 真正执行剪辑——所以你根本不需要 DaVinci Resolve 或 Adobe Premiere，我们全程 headless，它不会去启动别的剪辑软件、接管你的电脑，一切都在幕后完成；yt-dlp 是一个脚本，允许它从网上下载视频——YouTube、Instagram 或它找到的任何视频，都能拿来当 B-roll。保存指令，然后在左边 Customize 下面，确认装上 VidIQ。这是我这整套东西的「秘密武器」。搜索一下，VidIQ 就在这里，点进去、连接到 Claude；弹窗里点 Continue Connecting，跳转到 VidIQ 授权你的账号。VidIQ 慷慨地赞助了这期视频，但我要说明：早在他们同意赞助之前，我就一直在把 VidIQ 用于这套流程。我告诉 VidIQ「我要做一期讲我怎么剪视频的视频」，他们说「那让我们来赞助吧」，事情就这么自然。他们还设了一个页面：首月只要 1 美元——到 vidiq.com/matwolf（另一处作 matthewolf，未核实）就能用 1 美元用上你在全片会看到的 VidIQ MCP。电视端观众可以扫屏幕上的二维码。再次感谢 VidIQ 赞助。

[600s - 720s] 还在 Customize 标签页时，把 Hyperframes 也装上——它会负责视频里的一些动效。选中、连接到 Claude，浏览器里确认一下就连好了。回到我们的短视频剪辑项目。起步有两种方式：如果你已有 YouTube 或 Instagram 账号，可以让 VidIQ 去分析哪些视频表现好，让这套工具多产出类似的——我就是这么干的；如果没有，你可以给它几个你特别喜欢、想借鉴的 YouTube 频道做例子。既然我有频道，我就可以说：「用 VidIQ 分析我频道（@Mr. Eflows）上的短视频。看表现最好的 shorts，搞清楚它们为什么表现好——从脚本和画面两端都看。然后创建一个技能：生成同款、大概率会火的脚本；再创建一个技能：把视频剪成我现有视频的样子。剪辑技能可以用 FFmpeg 剪辑，也可以用 Hyperframes 和 Blender 增加视觉吸引力与质感。理想情况是：我给你一个我觉得有意思的东西的 URL，你写出大概率表现好的脚本；我照脚本录一段口播 A-roll 传回给你；你把这条视频剪成我现有视频的样子——切掉停顿、切掉口误、加 B-roll、加动效、加字幕、加背景音乐和音效。可以用 ElevenLabs 做背景音乐和音效。」另外我交代剪辑时有一个我喜欢加的点：「剪辑和初剪时，同时使用转写文本和波形图。」原因是：有时它读转写以为该切了，却会把你的单词切一半；所以它必须同时看波形，确认是在声音结束后下刀，而不是在它以为那个词结束的地方。提交，它就开始造这两个技能了，要跑一会儿。首先它会加载 VidIQ，对我现有视频做大量分析。为了加速，我把下面 Manual 切到 Auto 模式，让它一直跑；再点 Always Allow。可以看到它打开了 VidIQ，正在看我的视频，找出了离群值：这条 300 万播放、这条 190 万、这条 100 万——它在分析这些视频为什么在我频道上跑得这么好。

[720s - 840s] 这个过程我要提醒：它会是一个高度迭代的过程。它会回来交出两个技能——脚本技能和剪辑技能。脚本技能大概第一次就能相当到位；剪辑技能大概率会把 A-roll 剪砸，剪成我们不完全满意的样子。然后我们告诉它改哪里，改完再让它把改动更新进技能，就这样来回，直到剪辑越来越好。最终你会到达那一步：只说一句「去把这条剪了」，它完全知道该做什么。好，这一步跑了大约 15 分钟。它看了我表现最好的视频，创建了两个技能：short script 和 short editor。我把两个技能都保存，然后看看它做了什么。它分析了 shorts 并起草了双技能。这里是「你的 shorts 为什么能打」——这是它从所有 VidIQ 数据里读出来的：实体机器做出惊人动作的东西，比如飞行汽车、战斗机器人、身体扫描仪；有效的钩子是「This is the coolest thing I saw today」——这是我系列的名字，它知道我每条都以这句开场；人们会真的争论「看到的是真的还是 AI」的视频（我只发真实内容、除非另有说明，所以有点好笑：一旦人们要犹豫，说明我大概真找到了一件很酷的科技）；它还研究了我的节奏与视觉风格：与 B-roll 上层画面五五开、品红色字幕、B-roll 每 1.8 到 3.6 秒切换一次。它还需要我几样东西才能跑通整个流程：「需要你提供：连接一个文件夹」——放视频文件用的；「提供 ElevenLabs 的 API key」；还要我录一段测试录音。这里还有更多分析：VidIQ 的视觉分析把特斯拉和保镖片段描述为「CGI」，而脚本把它们呈现为真实演示——再说一次，「感觉像 AI 但其实是真的」的视频似乎表现相当好。它要我做的第一件事是连接文件夹：点下面这个小加号、Add Folder，我建了个叫 Claude Editor 的文件夹，打开它、发过去，让它知道我们用这个文件夹；这台设备上这个文件夹选「不再询问」、允许。现在它要我给 ElevenLabs 开网络访问：到 Customize，搜索 Eleven Labs，选中、Connect Claude，浏览器里出现连接页，点 Authorize。

[840s - 960s] 它会问装基础版还是完整版——装完整版，那个才有音乐、音效这些东西。继续，登录 ElevenLabs 账号，点 Add、再连接、全部授权，就连好了。

[960s - 1080s] 文件夹加了、ElevenLabs 连了，现在它要一段测试录音。它说「一条带脚本的原始 A-roll 就能看出剪辑在你真实人声上的效果」，但我不给脚本，因为我希望它在没有脚本时自己去拿。所以我告诉它：「我给你视频时，用 ElevenLabs 拿带时间戳的脚本。」它就去更新了技能。我点 Update、Update、Update——不知道为什么要点三下（笑）——现在技能更新完了：知道只要我丢一条没有转写的原始视频，它就会用 ElevenLabs 去抓转写。回到主项目文件夹，在 Context 下面把刚才建的文件夹加上，让它永远知道可以用那个文件夹；这个我叫 Claude Editor，打开、Always Allow。现在理论上全套就绪：我给它任何东西的链接，它就会给我写脚本；我照脚本录 A-roll、传回去，它就把整条剪出来。当然第一次大概不会完美——没人完美——但第一次就会相当接近，然后我们只要给反馈。快速找个选题：我最近看到 NVIDIA 的 Motion Bricks，觉得很适合做 short，因为视觉素材很酷、适合抓注意力。理论上我只需要：复制这个 URL，打开 Claude、进 Shortform Video Editor 项目、开个新窗口，把 URL 粘贴进去——

[1080s - 1200s] 可以看到它在读页面、加载了我们的技能 Mr. Eflow Short Script（转写存疑）——用的就是那个脚本技能；它在读 archive.org 的页面；现在又加载 VidIQ，在 VidIQ 上检索相似视频；它在给这段动画的视觉吸引力打分——真的在翻页面内容，找出视觉上最有吸引力的部分，留给下一阶段当 B-roll。这是它交回的东西：视觉 4/5 分；时长约 50 秒；一览要点：「对 AI 不感冒的人也会停下来看」这条没过——也就是说对不关心 AI 的人群不够强；但素材好、有天然争议、关联到人们关心的事（NVIDIA、游戏、Unreal Engine 等等）。它还基于 VidIQ 和它能访问的一切给了预估播放区间：你频道上游戏/3D 世界类 shorts 大约落在 1.5 万到 3 万播放，这条预计差不多——不是飞行汽车，但应该不差。然后给了我脚本。这条脚本没遵循我另一个「This is the coolest thing I saw today」开场的公式，因为这个技能还没被教过那种写法；这次我也不打算教——这是个全新的脚本设计师，我主账号那个技能已经完全按我的口味写了。现在我要做的是读一遍脚本：觉得行，就去录 A-roll；觉得不行，给反馈。老实说这条脚本相当扎实，通常会有「不只是 X，更是 Y」那种 AI 腔，这条没有，不像典型的 AI 脚本——它分析我过往脚本后写出了很像我的东西。

[1200s - 1320s] 它还给了标题建议：「这里的每一个动作都是 AI 生成的」「AI 正在实时给这个游戏做动画」「NVIDIA 刚刚取代了游戏动画师」；还做了 B-roll 镜头清单，这样进入下一阶段时它知道抓什么素材、放哪里。下一步就是去把这段录成 A-roll——我对着摄像机讲，basically 照脚本念，稍微即兴一点。我去录了。原始素材长这样：「看它走、慢跑；看它走、慢跑，然后跑起来；看它走、慢跑，然后全力跑起来。」可以看到，停顿、错误、重来全在里面。下一步就很酷了：把刚录的这个视频文件拖进来，连提示词都不用给，就这么提交。可以看到它加载了我们的技能 Mr. Eflow Short Editor：转写粗剪、拿 B-roll 和视觉素材、音乐、渲染、审片——这些它现在全会替我们做，我走开等着就行。好，我碰到了一个小问题：它在剪完之前停住了，因为几个权限拦截。如果你也遇到，这么修：它说 ElevenLabs 不接受我的视频去做转写，而且技能不允许别的工具来转写——那我们就告诉它「如果 ElevenLabs 转写不可用，就换一个备选工具，请更新技能」。它又说这台电脑上从 NVIDIA 站点下载被拦截——读不了 NVIDIA 的网站。这个问题应该由我去 Settings、Capabilities、勾选 Allow all sites 来解决，这样我让它读的所有站点都能读了，NVIDIA 应该就可以了。技能我也更新了，确认这些修改都写进去了，这样 ElevenLabs 挂了它也不会停。

[1320s - 1440s] 但它又提示我需要开一个新会话才能访问那些站点——它把该在新会话里发的话都给我了：「把这条剪进 Motion Bricks 的 short 里」，后面还带着脚本所在的 URL，我把这些全复制，开一个新会话，把视频文件拖进新对话，希望这次权限都正常了。结果又出了最后一个小问题：还是说这些 URL 被拦。我以为我已经全放行了，但没搞定。进 Settings→Capabilities，把 Create and Edit Files 打开；把「Use the Internet while making files」打开，确保它是开的；然后在 Allowed Websites 点 Manage——被拦的 URL 就填到这里放行。看到 URL 被拦，就到我刚才说的那个界面去放行。弄完之后，我 literally 只发了一句「还是不行」（笑），它就跑去把视频做出来了。我还没看过，这是它第一次尝试的成果：「这个角色的每一个动作都是 AI 实时生成的。这是 NVIDIA 的 Motion Bricks。看它走、慢跑，然后全力跑起来；它跃过长椅、坐下、捡起一把剑——而这些衔接没有一帧是手工动画。」不算差，但确实有问题：没做背景音乐；人声听起来有点怪，像被加了多余的压缩；有一处口误没剪掉；我还想让说话的间隙更紧。所以我给它一串反馈：把背景音乐搞定；我的人声有奇怪的 EQ 或压缩感，修一下；把我说话的间隙收紧；还留了几处说错的地方，修掉。另外我补了几条指令：务必同时用转写和波形，确保不会把我的词切一半；用 Hyperframes 加些动效——它一个额外动效都没加。最后我说：确保这些改动都更新进技能，以后所有视频也生效。让它带着这些反馈再跑一轮。

[1440s - 1560s] 好，大约 36 分钟后有了新版本。它其实还在跑，不过现在只是在更新技能文件，好让它以后知道怎么做这些事。可以看到它甚至创建了一些 Python 脚本来更好地完成这些目标：一个 short edit 的 Python 脚本，一个 GFX Kit 的 Python 脚本。来看新版：「这个角色的每一个动作都是 AI 实时生成的。这是 NVIDIA 的 Motion Bricks。看它走、慢跑，然后全力跑起来；它跃过长椅、坐下、捡起一把剑——而这些衔接没有一帧是手工动画。通常，游戏动画师需要手工构建这里的每一次过渡。」这版好多了：人声正常了、有背景音乐、加了更多图形与动效叠加、收紧了口播间隙、剪掉了我第一版里听到的错误。视频剪辑就是这么做的：不停地来回、来回，可以看到它基于这轮反馈更新了技能，我确认一句「好，确保从今往后都写进技能」。但你还能更进一步：接入 Blender 让它加 3D 物体。我切回另一个账号分享几个额外技巧。在我做的这条「Hynomas 世界模型」（转写存疑）视频里，快进预览，你可以看到一个实例：它跳进 Blender，创建了一个 GPU 的动画，然后用 Hyperframes 在上面加了文字。所以如果某个概念很酷但网上没有好素材，就让它用 Blender 生成图形并做动画，它会真的把它做进 B-roll 里。最近我还看到 Monique Price（我记得是给 NetworkChuck 做剪辑的）分享：[1560s - 1680s] 她把脑海里的一个分镜序列给了 Claude，它就做出来了。看这条视频，那个马桶大概率是 Blender 里建的，上面的文字和动画要么是 Hyperframes、要么也是 Blender 做的。这说明只要用提示词，你已经开始能把这种质量的动效放进自己的视频里。别误会：要精确到你要的样子，会有大量来回和迭代，但你能想到的大多数东西它都做得到。

另一件我喜欢在我的短视频项目里做的事：告诉它「回顾我最近十条 YouTube Shorts，找出表现异常好的离群视频，基于你的发现给我一份清单：如何改进我的脚本技能和剪辑技能」。我甚至不用告诉它用什么工具，默认它就会去调 VidIQ MCP——可以看到它已经把 VidIQ 拉进来了。点开小箭头能看到它的分步过程：用 VidIQ 拉取我频道的全部视频、看频道分析、拉出它认定表现最好的那些视频的转写、真的把短视频重新「看」了一遍。我知道我说过这期是 VidIQ 赞助的，但如果用 Claude 剪视频，它真的是最强大的工具之一。它还会读视频评论——观众给的反馈也会被拉进来；它能看视频的留存曲线。这套练习的思路，是几乎在造一个强化学习回路，让你的视频一次比一次好。大概每月跑一次：让它看上个月的所有视频、找表现出色的离群值、用这些信息更新技能，然后你未来的每一条视频都会因为这些发现稍微变好一点。每个月都这样做，你的视频只会越来越好：脚本更好、剪辑更好，技能也会持续学会自我改进，因为这个「让视频变好」的循环你在反复地跑。

[1680s - 1800s] 这是它给的反馈：开场帧——有的以一辆腾空的汽车开场、有的以机器人跳上车开场、有的以一张人脸特写开场；Genie 4 那条在第 2 秒切到「输入提示词」的画面；它告诉我们观众在第几秒流失，这样我们能针对性改进某些段落。脚本建议：第 2 句要说画面里物理上正在发生什么；中段埋词；用「看，它正在做 X」这样的句式；资历与参数放到标题卡上；但书限定到一个从句、或者干脆删掉。剪辑建议：第一帧永远应该是「实物真实在动」的画面。它还给了一批标题建议。可以看到我们拿到了一大堆能回喂给技能的反馈，未来的视频就会带着这些反馈去产出、 ideally 越来越好的结果。

最后谈谈我对这整套流程的最终想法：我的感受非常复杂。我自己做了 17 年视频剪辑——既剪自己的片子，早年也剪客户的。但说到这种「口播+信息」型短视频，我觉得 AI 现在已经剪得相当好了。长视频我仍然喜欢有个剪辑师、自己也亲手做一些精修，用老办法一点点抠到我满意的样子。但这类 short，我觉得完全可以快速批量出片，让 AI 替你干掉剪辑的大头。我之所以感受复杂，是因为我真的认为短视频剪辑师会越来越难接到活。

[1800s - 1920s] 拥抱这套东西的人也许能干五倍的活，这对他们很好；但也会有很多剪辑师失业。而这件事更大的含义是：我们会看到海量的内容洪水。创作、尤其是短视频创作的门槛已经低到任何人都能不停地量产视频。我不喜欢这样。但话一出口，又觉得这几乎像是在搞行业守门——毕竟我靠内容吃了 17 年饭，我不知道该怎么面对这一切。我这几天的大量时间都花在找有意思的信息、读文章、学新科技、找值得分享的东西上，我觉得我的价值越来越来自这里，尤其对短视频而言。我在 shorts 上并不是想剪赢别人，而是想做一个持续发现好东西、把它们摆到大家面前的人——那些大家本来可能看不到的东西。但我还是决定把这套东西整理出来公开。如果大家拿去用、这类视频越来越多，只会倒逼我把自己往上再抬一层、找到新的脱颖而出方式，因为这类产出正在被商品化。所以我恨自己正在成为商品化的推手之一，但这件事反正都会发生，不如把我知道的分享出来。

再次感谢 VidIQ 赞助这期视频。这个赞助来得非常自然：我告诉他们我在做一期「我怎么剪视频」的视频，而 VidIQ 是其中重要一环，他们说「那让我们赞助吧」，我说「好啊」——因为不管他们赞不赞助，我都会讲到 VidIQ。再次感谢。到 vidiq.com/matthewolf 就能以 1 美元用上 VidIQ，或者扫屏幕上应该出现的二维码。

[1920s - 1964s] 差不多就这些。这期比我预想的长，但我想做成一份详尽指南，而且路上碰到的每个小路障我都想指出来，这样你碰到时也能跨过去。最后提醒一句：这一切用 OpenAI 的 Codex 也能做，我做过；只是最近讲 OpenAI 的视频太多，这次想换 Claude 玩一遍。两边都行，你已经有哪个订阅就用哪个，不用再买别的。感谢你陪我一起 nerd，我真的很感激。我们下期见，拜拜。