Peter Yang Releases Open-Source Tool to Automate Short-Form Video Cuts

I Fully Automated My Shorts Editing (Get My Free Skill) — Peter Yang

Peter Yang

AI cannot produce polished short-form video in a single prompt, requiring iterative human review loops and automated self-evaluation rubrics to replace manual post-production.

AI-assisted summary of I Fully Automated My Shorts Editing (Get My Free Skill) — original source. Check the original for full context.

Source published:

Key ideas

  • Peter Yang has released `/video-short`, an open-source AI skill that automates turning raw talking-head footage into vertical shorts complete with B-roll, motion graphics, captions, and generative media. (0:00)
  • The local editing pipeline combines open-source CLI tools FFmpeg, yt-dlp, and Hyperframes, with optional cloud integrations for Google Veo video generation and Lyria music scoring via the Gemini API. (1:31)
  • Rather than relying on one-shot generation, the workflow introduces an interactive browser-based review tool where creators provide frame-by-frame visual and framing feedback back to the agent. (7:43)
  • The skill enforces mobile retention rules inside skill.md—such as a 3-second hook rule and a ban on low-contrast UI captures—and executes an automated self-evaluation pass before presenting edits. (11:34)

Detailed notes

## Background Published on YouTube on October 21, 2026, by creator Peter Yang, this video tutorial introduces his free, open-source `/video-short` skill designed to automate vertical short-form video production. Sponsored by Google, the video explains how creators can leverage frontier AI models within harnesses like Antigravity to transform raw, unedited footage into dynamic shorts complete with captions, B-roll, motion graphics, and background music.

## TL;DR Peter Yang demonstrates a four-step pipeline using his `/video-short` skill, which pairs open-source utilities like FFmpeg, yt-dlp, and Hyperframes with Gemini APIs to streamline short-form video editing. The workflow automatically drafts an edit plan, cuts verbal filler and awkward pauses, curates B-roll, and spins up a dedicated local web player where creators provide frame-by-frame feedback on framing and motion graphics. Built-in rules and self-evaluations inside the skill enforce mobile retention principles, such as securing a visual hook within the opening three seconds. Rather than promising one-click magic, Yang highlights an iterative review loop that delegates tedious post-production tasks so creators can focus on ideation and storytelling.

## Atmosphere & style The presentation is pragmatic, transparent, and grounded in real-world creative workflows rather than breathless AI hype. Yang speaks in a calm, conversational cadence, openly pointing out his own verbal tics, framing mistakes, and the reality that AI requires iterative human guidance rather than a flawless one-shot prompt. Watching him interact with the custom review UI feels like looking over the shoulder of an experienced product builder testing practical automations in real time.

## Bulletin - [00:00](https://www.youtube.com/watch?v=NWll-9p84b4&t=0s) Peter Yang has released `/video-short`, an open-source AI skill that automates turning raw talking-head footage into vertical shorts complete with B-roll, motion graphics, captions, and generative media. - [01:31](https://www.youtube.com/watch?v=NWll-9p84b4&t=91s) The local editing pipeline combines open-source CLI tools FFmpeg, yt-dlp, and Hyperframes, with optional cloud integrations for Google Veo video generation and Lyria music scoring via the Gemini API. - [07:43](https://www.youtube.com/watch?v=NWll-9p84b4&t=463s) Rather than relying on one-shot generation, the workflow introduces an interactive browser-based review tool where creators provide frame-by-frame visual and framing feedback back to the agent. - [11:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=694s) The skill enforces mobile retention rules inside skill.md—such as a 3-second hook rule and a ban on low-contrast UI captures—and executes an automated self-evaluation pass before presenting edits.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=NWll-9p84b4&t=0s) Short-form video has become an essential distribution format, with YouTube Shorts surpassing 200 billion daily views and platforms offering cash incentives for creators. - [00:00](https://www.youtube.com/watch?v=NWll-9p84b4&t=0s) The `/video-short` skill operates in four stages: installing dependencies, analyzing raw video for an edit plan, refining scenes via interactive feedback, and rendering the final cut. - [01:31](https://www.youtube.com/watch?v=NWll-9p84b4&t=91s) The pipeline requires three free open-source tools: FFmpeg for cutting and encoding, yt-dlp for downloading web B-roll, and Hyperframes for transcription and HTML motion graphics. - [01:31](https://www.youtube.com/watch?v=NWll-9p84b4&t=91s) Generative assets like Veo video clips and Lyria disco-house background scores can be integrated for fractions of a cent per call using a Gemini API key. - [01:31](https://www.youtube.com/watch?v=NWll-9p84b4&t=91s) To prevent accidental leakage to model training loops, users must never paste API keys into conversational chat windows, placing them directly into `.env` files instead. - [04:14](https://www.youtube.com/watch?v=NWll-9p84b4&t=254s) The workflow automatically detects narrative structure from a simple file path, generating an edit plan spanning full-screen hooks, B-roll overlays, and split-screen layouts. - [05:48](https://www.youtube.com/watch?v=NWll-9p84b4&t=348s) A built-in speech sanitization prompt removes pauses longer than one second, stutters, false starts, and repetitive filler words without introducing jarring jump cuts. - [07:43](https://www.youtube.com/watch?v=NWll-9p84b4&t=463s) An interactive web review player allows creators to leave timestamped notes on cropping, zoom levels, and UI clutter, which the AI agent automatically parses to re-render scenes. - [10:28](https://www.youtube.com/watch?v=NWll-9p84b4&t=628s) AI video editing is fundamentally an iterative feedback process rather than a one-shot generation task. - [10:28](https://www.youtube.com/watch?v=NWll-9p84b4&t=628s) The skill produces four distinct framing types: direct A-roll, tight punch-in crops for emphasis, split-screen layouts, and full-screen B-roll cutaways. - [11:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=694s) Mobile-first retention rules in `skill.md` mandate landing visual hooks within the first three seconds and explicitly forbid unreadable, low-contrast smartphone UI captures. - [11:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=694s) An automated evaluation rubric embedded in the skill forces the AI model to grade its own draft against retention criteria before presenting output to the creator. - [12:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=754s) Democratizing editing automation allows creators without professional video editing backgrounds to redirect their effort toward storytelling and ideation.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=NWll-9p84b4&t=0s) Overview of the `/video-short` Skill

Peter Yang launches his free `/video-short` skill, designed to transform a raw, unedited recording into a fully finished short-form video complete with B-roll, motion graphics, automated captions, and AI-generated video assets. To demonstrate the impact, he showcases side-by-side comparison clips highlighting the transition from a flat, single-take webcam recording to a dynamic, multi-layered vertical video.

Explaining the motivation behind building the workflow, Yang emphasizes the shifting media landscape across major networks: “Making great shorts like this is quickly becoming a core marketing skill because every social and video platform wants short video.” He notes that YouTube Shorts now generates over 200 billion views every single day, while platforms like X are actively offering thousands of dollars to incentivize short-form video creation. The entire creation workflow is structured into four distinct steps: installing the skill for free, uploading the raw video to receive an initial AI draft, reviewing and giving frame-by-frame feedback, and receiving the final rendered short. The video tutorial is sponsored by Gemini, and Yang demonstrates the workflow using Gemini 1.5 Pro in Antigravity, though he clarifies that any frontier AI model can work with the skill.

[01:31](https://www.youtube.com/watch?v=NWll-9p84b4&t=91s) Step 1: Installing the Skill and Configuring Dependencies

To begin, Yang demonstrates how to install the tool inside Antigravity by running a single command: `install the video-short skill globally from this URL github.com/petergyang/video-short`. When launched for the first time, the skill inspects the user's environment, detects any missing packages, and prompts to install them automatically. The local pipeline relies on three open-source tools: FFmpeg to execute video cutting and re-encoding, `yt-dlp` to query and download relevant B-roll footage directly from YouTube and X, and Hyperframes to transcribe spoken audio and generate motion graphic overlays.

```mermaid flowchart TD accTitle: Video Short Skill Pipeline Setup accDescr: Diagram showing the installation flow, core local dependencies, and optional Gemini API integration.

subgraph Installation["1. Installation & Environment"] A["GitHub Repository<br/>github.com/petergyang/video-short"] --> B["Antigravity Shell<br/>Gemini 1.5 Pro"] end

subgraph Dependencies["2. Core Dependencies"] B --> C["FFmpeg<br/>Cut & encode video"] B --> D["yt-dlp<br/>Fetch online B-roll"] B --> E["Hyperframes<br/>Transcribe & motion graphics"] end

subgraph OptionalAPI["3. External Cloud APIs"] B --> F["Google AI Studio<br/>Gemini API Key"] F --> G[".env File Configuration"] G --> H["Imagen 3 Video / Veo / MusicFX"] end ```

Beyond the free local dependencies, the skill optionally integrates with a Gemini API key to produce AI-generated imagery, generate original video clips with Veo, and score the video with custom music. Setting this up requires generating a key inside Google AI Studio and enabling billing, though Yang notes that running these models only costs a few cents or fractions of a cent per generation. Yang issues an explicit operational warning about secret hygiene: “Now, the important thing about the API key is when you go into Antigravity or another harness, you never want to actually paste the API key directly into chat here.” Instead, users should instruct the harness to open the project's `.env` file and paste the credentials directly into the marked variable slot to avoid inadvertently exposing API secrets to model training loops.

- Required open-source dependencies: FFmpeg, `yt-dlp`, Hyperframes. - Key generation destination: Google AI Studio (API Keys section). - Security requirement: Request the harness to open the `.env` file instead of passing keys in conversational chat.

[04:14](https://www.youtube.com/watch?v=NWll-9p84b4&t=254s) Step 2: Uploading Raw Footage and Generating an Edit Plan

Moving to the second step, Yang inputs a simple 54-second clip of himself speaking directly into the camera, saved as `raw.mp4` in his downloads directory. To initiate the workflow, he prompts the model: `/video-short`, tells it to locate the video file, and instructs it to edit the clip into a short that is as visually impressive as possible using motion graphics, web-sourced B-roll, AI video generation via Veo, and a Daft Punk "Get Lucky"-style funky disco house music bed generated via Lyria.

Rather than jumping blindly into an immediate render, the skill analyzes the source footage, detects where to make cuts, and drafts an initial scene-by-scene structural plan: - Scene 1: Speaker full screen to deliver the opening hook. - Scene 2: Cut to relevant B-roll footage. - Scene 3: Split-screen view featuring a graphic alongside the speaker. - Automated Sourcing: Curates source clips automatically, including clips from official product announcements such as the Veo launch blog and the YouTube Shorts blog.

Yang highlights that detailed initial instructions are entirely optional; users can merely supply the file path and let the model construct its own editing sequence, presenting all candidate B-roll links for manual inspection before proceeding.

[05:48](https://www.youtube.com/watch?v=NWll-9p84b4&t=348s) Automating Cuts for Pauses, Filler Words, and Retakes

Yang detours briefly into speech editing, explaining that while a concise 45-second short needs fewer structural trims, cleaning long-form raw footage is a massive editorial bottleneck. He shares a custom, reusable prompt built into the `/video-short` tool that automatically sanitizes spoken delivery without producing jarring edits.

The prompt executes specific editorial rules: it truncates pauses to one second or less, strikes through repetitive phrases, removes false starts, clears restated takes, and cleans up stutters and mid-sentence self-corrections. It also targets verbal tics: Yang confesses, “For example, I like to say 'all right' and 'so' a lot. And I also like to say 'right?' at the end of every line, which is very annoying to listen to.” Crucially, the prompt directs the engine to preserve pacing by cutting mistakes cleanly without producing a jarring series of choppy jump cuts. A baseline version of this silence- and retake-pruning system is embedded natively into the skill file to prepare clean spoken tracks automatically.

[07:43](https://www.youtube.com/watch?v=NWll-9p84b4&t=463s) Step 3: Frame-by-Frame Scene Review and Iteration

Once the engine processes the plan and identifies any billable generation tasks (such as Veo clips or music generation), it compiles the first edit cut and automatically spins up an interactive browser-based review tool. This visual interface plays back the edited short, complete with light and dark mode toggles, and lets the creator leave granular, timestamped notes on individual scenes or submit global video notes.

```mermaid flowchart LR accTitle: Frame-by-Frame AI Review Loop accDescr: The feedback iteration cycle between the web review UI and Antigravity.

A["Raw Edit Generated"] --> B["Local Web Review UI"] B --> C["Review Scene by Scene<br/>Pin Comments & Framing Fixes"] C --> D["Submit Batch Notes"] D --> E["Antigravity / Gemini Agent<br/>Re-analyzes Framing & Visuals"] E --> F["Updated Video Cut"] F --> B ```

During his live pass through the review player, Yang logs specific visual adjustments: - Head Framing: In the opening scene, the vertical crop cuts off the top of his head, so he leaves a comment requesting a slight zoom-out. - B-Roll Integration: He confirms that the Veo-generated video snippet and a dynamic Hyperframes-sourced collage overlaid with a bold "200 billion" graphic rendered correctly. - Layout Simplification: In a scene showcasing an early mockup of the review interface, he enters a comment asking to eliminate an unnecessary notes section to reduce visual clutter: “Can you remove the overall note section of the review page to simplify it?” - Split-View Alignment: In the bottom half of a split screen, his head is cut off again, prompting another note to adjust framing.

Yang clicks "Send Notes," prompting Antigravity to parse the instructions. The AI agent inspects frame grabs, verifies head positions, shifts cropping coordinates downwards, and regenerates the sequence. When Yang reopens the interface, the head framing errors are resolved and the requested visual elements are cleaned up, leaving the project ready for final rendering.

[10:28](https://www.youtube.com/watch?v=NWll-9p84b4&t=628s) What the Video Short Skill Generates

Peter clarifies realistic expectations about what AI video editing currently looks like: it is not magic that happens in a single prompt. Even reaching the baseline edit shown earlier took multiple iterative rounds of human feedback, rather than popping out fully formed on the first attempt. As Peter notes, “So, this isn't a oneshot scale to turn a raw recording into a polish short.” However, pairing an interactive review interface with programmable assets like HTML motion graphics radically simplifies and accelerates what used to be a tedious manual editing timeline.

To summarize the tool's core output, the skill analyzes the source footage and splits the timeline across several distinct shot types: standard A-roll (direct speaking to camera), dynamic punch-ins (a tighter camera crop for emphasis), split-screen layouts (the speaker on the bottom half with supporting footage above), and full-screen B-roll takeovers. Beyond raw cutting, the skill orchestrates automated subtitle captions, custom motion graphics rendered directly via HTML, web-sourced reference footage, and generative AI assets—specifically AI video powered by Google Veo and background music generated via Lyria, both enabled through a Gemini API key.

```mermaid flowchart TD accTitle: Video Short Layout and Asset Pipeline accDescr: Pipeline showing how raw footage is split into layouts and enriched with visual and audio assets.

subgraph Cuts ["Framing & Cut Types"] ARoll["A-Roll (Direct to Camera)"] Punch["Punch-In (Tight Crop)"] Split["Split Screen (Speaker Bottom / B-Roll Top)"] FullB["Full-Screen B-Roll"] end

subgraph Assets ["Enrichment Layer"] Caps["Captions"] HTMLMotion["HTML Motion Graphics"] WebBRoll["Web B-Roll Clips"] GenAI["AI Video (Veo) & Music (Lyria)"] end

Cuts --> Assets ```

- **Framing Variations**: Cuts footage into A-roll, punch-ins, split-screens, and full-screen B-roll cutaways. - **Graphic Overlays**: Writes dynamic on-screen motion graphics using standard HTML code. - **Generative Media**: Calls the Gemini API to produce generative video using Google Veo and background scores using Lyria.

---

[11:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=694s) Editing Rules and Built-in Evals Inside `skill.md`

Diving into the actual repository repository, Peter inspects the underlying `skill.md` file, which is packed with battle-tested short-form video principles. He credits fellow AI creator Nate Herk for teaching him many of the retention rules baked into the prompt instructions. The file instructs the model on strict mobile viewing habits, establishing that the first three seconds must immediately grab attention by pairing the speaker with a compelling visual element where the graphic visually demonstrates what the spoken words do not. It also commands the model to favor real footage over tiny, illegible elements, explicitly warning against dim UI panels or text details that disappear on small phone displays.

To ensure the model adheres to these editing standards without drifting, Peter built an automated self-evaluation step directly into the skill definition. Rather than handing the raw output straight back to the user, the model pauses to inspect its own draft against a rubric before presenting it. Peter explains, “And also I built a eval which is just best practice for setting up a scale where you have the AI actually check its own work before it presents the result to you.” This evaluation pass verifies concrete criteria, such as whether the opening hook lands cleanly within the first three seconds and whether each supplementary visual communicates new information beyond the voiceover.

- **The 3-Second Rule**: Hook viewers immediately within the opening three seconds by combining speaker presence with visual contrast. - **Visual Complementarity**: Every graphic or B-roll shot must show what the spoken script is not saying, rather than duplicating words. - **Mobile-First Clarity**: Prohibits low-contrast, dim UI screenshots or microscopic UI panels that cannot be parsed on a smartphone screen. - **Internal Evals Checklist**: The skill forces the AI to grade its own cut before displaying it, confirming pacing, visual storytelling, and hook timing.

---

[12:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=754s) Community Tools, Ecosystem Plugs, and Democratizing Video Creation

Peter wraps up the walkthrough by sharing additional tools available on his GitHub profile (`github.com/petergyang`). These include popular open-source projects such as "No AI Slop," a repository with over 12,000 GitHub stars designed to strip generic AI mannerisms from written prose, as well as "Human Review," an interactive review workflow resembling Google Docs that allows creators to give line-by-line feedback to language models. For users wanting deeper resources, he mentions his membership platform, Behind the Craft (`behindthecraft.com`), which hosts more than a dozen production-ready skills alongside software discounts for developer tools like Replit, Lovable, Bolt, and Groq.

Reflecting on the broader impact of AI-assisted editing, Peter is open about his own skill level, readily admitting that seasoned video editors may produce far slicker edits manually. Yet he emphasizes that lowering technical barriers is the entire objective of building these workflows. As Peter puts it, “Uh if AI can help us edit our shorts, then we can focus more on storytelling and making more and better content.” He invites creators to clone the repository from GitHub, test the skill on their own footage, and focus their energy on ideation and narrative craft.

- **"No AI Slop"**: Peter’s open-source repository with over 12,000 stars dedicated to eliminating AI clichés from written text. - **"Human Review"**: A browser-based interface enabling users to comment and provide targeted, line-level feedback to AI drafts. - **Behind the Craft**: A paid platform offering over twelve advanced skills and partner perks for AI tools like Replit, Lovable, and Bolt.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=NWll-9p84b4&t=0s) Making great shorts like this is quickly becoming a core marketing skill because every social and video platform wants short video. — Peter Yang

> [12:34](https://www.youtube.com/watch?v=NWll-9p84b4&t=754s) if AI can help us edit our shorts, then we can focus more on storytelling and making more and better content. — Peter Yang

## Mentioned Peter Yang, Nate Herk, Gemini, Gemini 1.5 Pro, Antigravity, FFmpeg, yt-dlp, Hyperframes, Google AI Studio, Veo, Lyria, YouTube Shorts, X, No AI Slop, Human Review, Behind the Craft, Replit, Lovable, Bolt, Groq

## Video description 📌 Get my free video short skill here: https://github.com/petergyang/video-short 🤝 Get a Gemini API key here: https://aistudio.google.com/?utm_source=peteryang&utm_medium=partner&utm_campaign=26Q3-peteryang_ais

AI can now turn a raw talking-head recording into a polished video short, complete with B-roll, motion graphics, captions, AI video, and music. In this tutorial, I walk through my free /video-short skill in 4 steps: install it, upload your raw recording, give AI feedback frame by frame, and render the final short. I use Gemini 4 Argon in Antigravity, but any frontier model works with the skill.

What I cover: (00:00) What my free /video-short skill can do (01:31) Step 1: Install the skill (04:14) Step 2: Upload your raw recording and get an edit plan (05:48) My prompt to cut pauses, retakes, and filler words (07:43) Step 3: Give AI feedback frame by frame (10:18) Step 4: Approve and render the final short (11:34) Video editing best practices inside the skill file

This video is sponsored by Google but opinions are my own.

Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/

Subscribe to this channel - more tutorials coming soon!

[Source](https://www.youtube.com/watch?v=NWll-9p84b4)

中文简报

Peter Yang 发布开源工具，实现短视频剪辑自动化 AI 无法通过单次提示词生成精修短视频，需要人机迭代审核循环和自动化自我评估标准来取代手动后期制作。 Peter Yang 展示了使用其 `/video-short` 技能的四步工作流，该方案将 FFmpeg、yt-dlp、Hyperframes 等开源实用程序与 Gemini API 相结合，以简化短视频剪辑流程。该工作流可自动拟定剪辑方案，去除口头填充词和尴尬停顿，筛选补充镜头（B-roll），并快速生成专用的本地 Web 播放器，供创作者对构图和动态图形提供逐帧反馈。技能内部内置的规则和自我评估机制严格执行移动端留存原则，例如确保在前三秒内出现视觉钩子。Yang 并没有宣称其具备一键生成的奇效，而是强调了迭代审核循环的作用，将繁琐的后期制作任务分工出去，以便创作者专注于创意构思和叙事。 Peter Yang 发布了开源 AI 技能 `/video-short`，可自动将未经剪辑的单人出镜原始素材制作成包含 B-roll、动态图形、字幕和生成式媒体的竖屏短视频。 (0:00) 该本地剪辑工作流结合了开源 CLI 工具 FFmpeg、yt-dlp 和 Hyperframes，并支持通过 Gemini API 可选集成 Google Veo 视频生成以及 Lyria 音乐配乐等云端功能。 (1:31) 该工作流并非依赖单次直接生成，而是引入了一个交互式基于浏览器的审核工具，供创作者向代理反馈针对画面和构图的逐帧视觉意见。 (7:43) 该技能在 skill.md 中强制执行移动端留存规则（如 3 秒视觉钩子原则和禁止低对比度 UI 抓图），并在展示剪辑结果前执行自动化自我评估审查。 (11:34)

精炼阅读摘要

Peter Yang 推出了名为 /video-short 的自动化工作流，通过结合开源工具链与前沿大语言模型，将未经剪辑的单机位原始视频转化为包含动态图形、网络 B-roll 和 AI 生成素材的竖屏短视频。 本地工具链与云端 API 的混合架构 (1:31) 该工作流运行于 Antigravity 等 AI 运行环境中，主要由三大开源工具驱动：利用 FFmpeg 执行视频剪切与重新编码，利用 yt-dlp 直接从 YouTube 和 X 等平台检索并抓取相关的 B-roll 补充素材，利用 Hyperframes 负责音频转录与生成 HTML 动态图形覆层。演示中工作流由 Gemini 赞助并使用 Gemini 1.5 Pro 进行操作，但理论上兼容其他前沿大模型。 除免费的本地开源依赖外，系统还支持通过接入 Google AI Studio 的 Gemini API 来扩展生成能力，包括调用 Veo 生成 AI 视频以及调用 Lyria 制作定制背景音乐。Yang 明确提出了密钥安全警告：用户绝不应将 API 密钥直接粘贴到大模型对话框中以防训练泄露，而必须指示智能体打开项目的 .env 文件，将凭据写入指定的变量配置中。 口语冗余清理与分镜自动化规划 (5:48) 在接收到原始视频后，系统会先对音频和画面进行分析，而不是盲目开始渲染。它会自动生成分镜规划，将时间线拆分为开场钩子（全屏主讲）、相关 B-roll 切换以及图文结合的分屏视图，并自动搜集官方博客等公开渠道的参考素材链接供用户核对。 为解决长素材剪辑中最耗时的修剪瓶颈，技能内置了专门的语音净化提示词：自动将超过 1 秒的停顿截断，删除重复语句、抢跑抢话、重录片段以及中途自我纠正的结巴。该提示词还专门针对高频口头禅进行剔除（例如 Yang 自述常说的“all right”、“so”以及句尾习惯性的“right?”），并在修整口误的同时保持语言节奏自然，避免产生生硬的跳剪。 基于网页界面的逐帧审查与迭代闭环 (7:43) Yang 强调当前的 AI 视频剪辑并非“单次提示词即可完成”的魔法，初稿呈现仍需多轮人机交互。系统初次剪辑完成后，会自动生成一个支持明暗模式切换的本地网页端审查工具。创作者可以在具体时间戳上留下精确到场景的修改意见，例如指正竖屏裁剪时头部上方被切断的构图问题，或是要求移除多余的面板以简化界面视觉。 当用户提交批注后，后端的 AI 智能体会自动解析指令，检查关键帧抓取图像，调整画面裁剪坐标，并重新生成对应片段。整个输出体系支持标准 A-roll 主讲画面、强调重点的局部推近特写（punch-ins）、上下分屏以及全屏 B-roll，并通过 HTML 动态图表丰富画面层次。 移动端留存法则与内置自我评估机制 (11:34) 技能底层文件 skill.md 固化了短视频的传播与留存规则（部分源自创作者 Nate Herk 的经验总结）。其中包括严格的“前 3 秒法则”，即开头必须通过主讲人出镜配合高反差视觉元素迅速吸引注意力；要求补充素材必须展示口播没有表达的信息，而非文字内容的简单画面复述；并明确禁止在手机小屏幕上展示低对比度、昏暗或字号过小的界面截图。 为了防止模型在执行过程中发生逻辑漂移，系统在最终呈现草稿前强制执行一道“自我评估”（Eval）工序。AI 必须依照预设的核对清单对自身生成的剪辑方案进行打分，核验前 3 秒钩子是否达标、补充画面是否具有增量信息，确认无误后方才输出最终视频草稿。 Yang 指出，资深专业剪辑师的手工精修效果可能依然优于自动化工具，但该项目的核心目的在于消除繁琐的技术摩擦，让创作者将时间集中在构思与叙事上。目前该流程仍依赖创作者通过交互界面进行多轮校验与微调，而非完全无人工干预的自动化。

Source quotes

[0:00] Peter Yang: Making great shorts like this is quickly becoming a core marketing skill because every social and video platform wants short video. [12:34] Peter Yang: if AI can help us edit our shorts, then we can focus more on storytelling and making more and better content.

Original transcript

[0:00] Today, I'm really excited to launch my [0:02] free video short scale to turn any raw [0:05] recording into a polished video short, [0:08] complete with B-roll, motion graphics, [0:11] captions, and even AI generated video. [0:15] Here are some quick clips that directly [0:17] compare a raw recording with a polished [0:20] short using my scale. [0:32] Making great shorts like this is quickly [0:35] becoming a core marketing skill because [0:38] every social and video platform wants [0:40] short video. YouTube shorts alone gets [0:43] over 200 billion views a day and X is [0:46] even offering to pay thousands of [0:48] dollars for creators to make shorts. So, [0:50] in this video, I'm going to show you [0:52] exactly how to use my free skill to turn [0:55] any recording into a polish short in [0:57] four steps. First, install the skill for [1:00] free. Second, upload your raw recording [1:04] and get a draft edit from AI. Third, [1:08] give feedback to AI frame by frame and [1:10] iterate. And fourth, get back a polish [1:14] short. Now, this video is sponsored by [1:16] Gemini, and I'm going to be using Gemini [1:19] 4 Argon because I generally think it's [1:21] the best multimodal model with great [1:24] APIs for image, video, and music [1:26] generation, but any Frontier model works [1:29] with my scale. Okay, so step one is just [1:33] to install the scale. Now, I'm using [1:35] anti-gravity here with Argon as my [1:38] model. To install a skill, all you got [1:40] to do is paste in install the videos [1:43] short scale globally from this URL [1:45] github.com/pergyang/videosshort. [1:50] Awesome. So, here's the install [1:51] instructions step by step on the re on [1:54] the GitHub for the skill. And the first [1:56] time you run the scale, it's going to [1:58] check if you have the following [1:59] dependencies installed and offer to [2:01] install what's missing. So, it needs [2:04] ffmpeg to cut and encode the video. It [2:08] needs yt-dlp [2:10] to find and pull b-roll clips from [2:13] YouTube x and other sites. And it also [2:16] needs hyperframes, which is a free [2:19] open-source tool to transcribe your raw [2:22] recording, add motion graphics, and [2:24] more. And optionally, it needs a Gemini [2:27] API key to generate images with nano [2:30] banana video clips with VO and music. [2:34] Now, everything here is free except for [2:36] the Gemini API key, but even that it [2:39] only costs a few cents or even a [2:41] fractions of a cent to generate some of [2:43] these clips. But let me quickly show you [2:45] now how to set up the API key. So, all [2:48] you got to do is go to Google AI Studio [2:50] and navigate to API keys. And you can [2:54] see all my API keys here. Just hit [2:56] create a new API key and it's going to [2:59] generate a new API key for you. And make [3:01] sure you also turn on billing for the [3:03] API key. So for example, I got four keys [3:05] here, but only this one has billing [3:07] turned on because like I said, it does [3:10] cost a few cents or a few fractions of a [3:12] cent to get Gemini to generate AI [3:16] videos, music, and so on. Now, the [3:18] important thing about the API key is [3:21] when you go into anti-gravity or another [3:24] harness, you never want to actually [3:27] paste the API key directly into chat [3:29] here. Instead, what you want to say is, [3:31] hey, can you open the EMV file so I can [3:35] paste in the latest API key? And then [3:38] what it's going to do is open a file and [3:40] it's going to be very clearly marked [3:42] where you should paste in the Gemini API [3:43] key. And the reason here, if you don't [3:45] understand, is because if you paste the [3:46] API key here in chat, maybe the AI will [3:49] learn about it and it'll be in training [3:52] and [laughter] you don't want your key [3:53] to leak. So definitely take the step for [3:56] security, ask it to open EM file and [3:58] paste the key there. [4:00] And of course, you don't actually need [4:01] to use a Gemini APA key if you don't [4:03] want to. I just think it's very useful [4:05] for adding music background to your [4:07] clip, for adding awesome animated AI [4:10] videos, and more that I'll show you [4:12] soon. All right, so step two is to [4:15] upload your raw recording and get a [4:18] draft edit from my skill. So, this is my [4:20] raw recording. Let me go ahead and play [4:21] it. It's basically just me talking to [4:23] the camera. There's nothing fancy here. [4:25] Just a 54 second clip of me talking. So, [4:29] uh, what I did next was I said /vide [4:32] short to trigger the scale. And here's [4:35] my raw recording raw.mpp4. Go find it in [4:37] my downloads folder. And then edit this [4:40] into a short that's visually impressive [4:42] as possible with motion graphics, [4:44] b-rolls you find online, AI video from [4:46] VO, generate a funky disco house music [4:50] with Laria, similar to Da Punk's Get [4:52] Lucky, [laughter] and so on so forth. [4:54] Now, you don't actually have to provide [4:55] all these instructions. You can just [4:57] say, "Hey, here's the raw recording. Go [4:59] ahead and edit this short." Because the [5:01] next step here is that AI is going to [5:02] come up with a plan. So, it's going to [5:04] look through my raw recording and it's [5:06] going to decide where it needs to cut [5:08] and then it's going to give a scene by [5:10] scene plan of how it wants to lay out [5:14] the short edit. So, for example, scene [5:17] one, it wants to do speaker full screen. [5:20] Scene two, it wants to use a B-roll. [5:22] Scene three wants to use a graphic in a [5:25] split view. and so on so forth. So, [5:26] here's the feedback and basically uh [5:30] here's all the B-roll sources that it [5:31] found. So, it found some great B-rolls [5:34] on VO from the VO announcement blog on [5:37] shorts from the YouTube shorts blog and [5:39] so on and so forth. And I can check each [5:40] of these B-roll sources to see if it's [5:42] actually clips that I want to use. After [5:44] this, you want to give feedback to AI. [5:48] And before we get into like the B- rolls [5:49] and the motion graphics and stuff, I [5:51] want to show you guys some something [5:52] else. So, in a video short, you know, [5:56] it's only 45 seconds or shorter. So, you [6:00] probably don't have to do too many cuts [6:01] for retakes and fewer words and stuff. [6:04] But for a longer video like this [6:05] tutorial, this is a prompt that [6:07] basically is a lifesaver for me. And [6:09] this prompt I came up with myself, but [6:12] basically what it does is it shortens [6:14] pauses to 1 second or less. It strikes [6:16] through repeat phases, false starts or [6:19] restate it takes, cleans up all the [6:21] stuttering that I do and [6:22] self-corrections. And it also strikes [6:24] through a bunch of filler words. So, for [6:26] example, I like to say all right and so [6:29] a lot. And I also like to say right at [6:32] the end of every line, which is very [6:33] annoying to listen to. So, I basically [6:35] just got AI to cut a bunch of the stuff. [6:37] And the last one is like cut the stuff [6:39] without making a whole bunch of choppy [6:41] jump cuts. Now again this is not as [6:43] relevant for shorts but for longer [6:45] videos I think it's very important to [6:46] build your own version of this and I've [6:49] also included a basic version of this [6:51] directly in the video short scale so [6:54] that if you make a short it can also cut [6:56] silences and cut retakes for you. You [6:58] know I want to make this as easy as [7:00] possible for you guys to actually do. [7:02] Okay so let's go back here and [7:03] anti-gravity also told me where it wants [7:06] to generate paid versions of you know v [7:09] clips and the music bed. this [snorts] [7:12] is important to review because it's [7:13] going to cost you I don't know a cent or [7:15] two to generate this stuff. But anyway, [7:18] I basically gave AI some feedback. So I [7:20] said, "Hey, for scene seven and scene [7:22] eight, I like the shorts collage from [7:25] this other clip that I found online from [7:26] hyperframes. Can you please use that and [7:30] put a big number 200 billion number [7:32] overlay on top of it to see how many [7:34] shorts people are consuming on YouTube?" [7:36] All right. So, I gave some feedback and [7:38] basically it made the first cut edit of [7:40] my short. Now, here's where the magic of [7:43] the scale is. Step three, which is to [7:45] review the clip and give feedback frame [7:48] by frame. So, what does that mean? As [7:51] part of the skill, once it generates the [7:53] first cut, it's going to open this [7:55] review page that I'm going to show you [7:57] right here. And this review page is [8:00] basically the AI polish short. And I can [8:03] actually leave feedback by scene. So [8:05] let's first of all play the short and [8:06] take a look. All right. So basically [8:08] it's going to play my actual edited clip [8:11] and I'm going to give it some feedback. [8:12] So for example here it is cutting off my [8:14] head. Like I want to zoom out a bit on [8:17] my head so it's not cut off. So I'm [8:20] going to add a comment here. And this [8:22] whole interface by the way is built into [8:23] the scale. And I even built like dark [8:25] and light mode here. Let's keep playing [8:27] here. [8:33] Yep. And this video clip, by the way, is [8:34] generated by VO. So it looks great. [8:41] And here is the zoom out of all the [8:43] clips from the hyperframes video that I [8:45] found. And you know, I gave it feedback [8:47] to overlay the 200 billion on top. And [8:49] looks like it did it. So far, this all [8:51] looks pretty good. I don't really have a [8:52] ton of feedback. It looks like it [8:54] actually locked up the review page here, [8:56] uh, similar to this, but mobile [8:58] optimized. I think maybe we can ask it [9:00] to get rid of this overall no thing cuz [9:01] it makes it too complicated. Let's just [9:03] say, can you remove the overall note [9:06] section of the review page to simplify [9:10] it? Right? I add a comment here. Okay. [9:12] And here again, I think this top part [9:14] looks good, but the bottom part is [9:15] cutting off my head. So again, zoom out [9:18] a bit on my head here, too, to avoid [9:21] cutting it off. All right. So, this [9:25] feedback is not super sophisticated. [9:27] There you go. So, here's my feedback. [9:28] And I can also add an overall note to [9:30] the video like maybe I want to color [9:31] grade it or do something else. Let's go [9:33] ahead and send our notes here. Okay, so [9:35] we sent our notes and let's go back to [9:37] anti-grab gravity. Awesome. So it's [9:39] taking into account our feedback. You [9:42] can see here it's checking head framing. [9:44] It's analyzing these images that my head [9:46] is cut off and hopefully it will fix [9:48] this stuff and open the review page [9:50] again. Let's skip ahead until that [9:52] happens. Okay, we're back and AI has [9:55] finished making it edit. So, let's take [9:56] a look at the review page again. So, [10:00] it's positioned my head to be a little [10:01] bit lower. So, top half is not cut off. [10:04] That looks good. And I think it was [10:06] scene 12 where it's removed the bottom [10:10] input box to make it a little more [10:11] clear. And scene 13 is also positioning [10:15] my head well. So, basically fixed [10:16] everything. And now I can say approve [10:18] and render final. and it will render the [10:21] final clip for me to upload to [10:24] Instagram, YouTube, and so on. And of [10:27] course, you can also use AI to upload [10:28] this stuff. Now, to be very clear, it [10:32] didn't do all this in one shot. It took [10:34] a few more rounds of feedback to even [10:36] get to the before version that you saw. [10:38] So, this isn't a oneshot scale to turn a [10:41] raw recording into a polish short. But I [10:44] think with, you know, interfaces like [10:46] this and the ability to add motion [10:48] graphics and more, it just makes it way [10:51] easier. Awesome. So before we wrap, I [10:54] just want to summarize again what the [10:55] skill can do. So first of all, it can [10:58] cut every short into a mix of a-roll, [11:02] which is me talking to the camera, a [11:04] punch, which is tighter shot, split, [11:06] which is me talking on the bottom half, [11:08] and a b-roll on the top half, and also [11:10] full screen bureau. In addition, it can [11:13] add captions, motion graphics written in [11:16] HTML to bring the short to life, b-roll [11:19] footage that you can find online, and AI [11:23] video and music generated with Gemini VO [11:26] and Liaria. I think that's how you [11:29] pronounce it. And this step, of course, [11:31] needs a Gemini API key as I walk through [11:33] before. And let's also take a look at [11:35] the actual skill file. So you can go in [11:37] here and hit this skill.md because [11:40] there's actually a lot of really good [11:41] best practices for making video shorts. [11:43] And here I have to give some credit to [11:45] Nate Herk, which is [clears throat] [11:46] another awesome AI YouTuber that taught [11:48] me a lot of these best practices. [11:50] So for example, when you are making a [11:52] short, you got to capture the viewers's [11:54] attention in the first 3 seconds with [11:57] the speaker and the strongest visual. So [11:59] the visual should show what the words [12:01] don't show. And then prioritize real [12:03] footage first. skip dim UI panels and [12:07] other things that are too tiny to see on [12:09] a mobile screen. There's a bunch of best [12:11] practices here. I encourage you to read [12:13] the scale file. And also I built a eval [12:17] which is just best practice for setting [12:18] up a scale where you have the AI [12:21] actually check its own work before it [12:23] presents the result to you. So here we [12:26] have a check of does the hook land in [12:28] the first three seconds? Does each [12:30] visual show something the words don't [12:31] say? And so on and so forth. Awesome. [12:34] Now, a quick plug. I have a few more [12:37] free popular open source skills on my [12:39] GitHub. It's just github.com/petergy [12:42] gang, including no AI slop to remove all [12:45] slop from your writing that has 12,000 [12:47] stars and human review, which is [12:49] interface very much like Google Docs and [12:51] what I just showed you. So, you can [12:53] provide AI feedback line by line. You [12:56] can also check out my paid membership [12:58] website at behindthecraft.com [13:01] for a dozen plus more skills that are at [13:04] the same quality level as my open- [13:07] source skills. It also comes with free [13:09] AI tool discounts to Grockbot, Replet, [13:11] lovable, and a lot more popular tools as [13:14] well. So, you can check it out at [13:16] behindthecraft.com. [13:18] Awesome. So once again, you can find my [13:20] video short skill in my GitHub and I've [13:23] also linked to it in the video [13:24] description and the first comment. Last [13:26] but not least, I want to acknowledge [13:28] that, you know, I'm not an expert in [13:30] making video shorts and maybe you can [13:31] make better shorts than what I just [13:32] showed you, but that's the whole point [13:34] of AI and democratizing access to this [13:37] stuff, right? Uh if AI can help us edit [13:39] our shorts, then we can focus more on [13:41] storytelling and making more and better [13:43] content. Now go out there and make more [13:47] amazing video shorts. See you soon.