chatgpt-video-editing-skills: Two Agent Skills That Keep a Short-Video Edit Verifiable
用 ChatGPT／Codex 安裝 AI 剪輯環境，並依八大步驟完成可驗證的短影音剪輯。
At a glance
- What is it?
- The repository splits environment setup from the actual edit, keeps the source file untouched, and will not render a 1080x1920 master until a 720p preview has been approved. Here is what that buys you, and where it costs you.
- Who is it for?
- Adopt it if you already have a compatible agent that can run shell commands, you are editing Traditional Chinese short video, and you want the source file protected and every stage approved before the next one starts. Do not adopt it if you want a one-click editor, if you refuse to send audio to ElevenLabs and cannot accept lower-confidence timestamps, or if you are not comfortable letting an agent install tools.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 54 days ago.
- What is it written in?
- Mainly Shell, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 17, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem: an agent that says "done" before anything is rendered
Most agent-driven editing attempts fail at the same place. The model prints an FFmpeg command, describes the cut it would make, and calls the job finished. Nothing was executed, or something was executed and never decoded back to check it. This repository is built around refusing that. Its README states the package treats actual output and QA evidence as the standard, and does not call a plan, a command, or an unverified file a finished product.
The audience is narrow and specific. You need an agent that can read and write local files and run terminal commands, and you need to be producing vertical short video with Traditional Chinese subtitles. The README is written in Traditional Chinese and the subtitle font it installs is the Source Han Sans TW subset, so the pipeline is aimed at that language. If you edit in English, the font step is dead weight.
The design decision worth noticing is the split into two skills. chatgpt-video-editing-setup checks, installs, repairs and verifies the environment. chatgpt-short-video-editor does transcription, strategy, rough cut, subtitles, preview, QA and final output. The README's own table gives the boundary: the setup skill does not upload footage, transcribe, edit or output video, and the editor skill does not silently install tools or produce a final master before preview approval. That separation is the whole argument of the project.
How the eight-step workflow actually sequences the work
The pipeline is a fixed order, and each step gates the next. Step one runs ffprobe against the source to confirm its specification and that it decodes, with the original file left unchanged. Step two asks for file-level consent before upload, then gets word-level timestamps from ElevenLabs Scribe v2. Step three is content work: finding the hook, the main line, what can be cut, and what still needs confirmation.
Step four is where the agent has to commit in prose. It proposes four to eight plain-language sentences describing the editing strategy and waits for approval before choosing cut points or creative elements. Step five builds an EDL on complete word boundaries, keeping 30 to 200 ms of boundary room and roughly 30 ms audio fades. Step six handles colour, cards and subtitles, with colour only when technically necessary or approved, static cards through Pillow, optional animation through HyperFrames, and subtitles composited last in Source Han Sans TW.
Steps seven and eight are the part most agent workflows skip. The agent renders and checks one complete 720p preview. Only after explicit approval of that preview does it output a 1080x1920 master, and the master still needs its own independent inspection and full decode before delivery. The README is blunt about why: a command is not an execution, a frame is not full QA, and a preview is not a delivery.
All new artifacts land in an edit/ directory beside the source, containing project.md, transcripts/, corrected-transcript.md, edl.json, master.srt, clips/, animations/, qa/, preview.mp4 and final.mp4. The source is never overwritten, moved, renamed or deleted.
Installing the skills with npx and running a first edit
The README's recommended path is the interactive installer. Run it from your project directory and it asks which skills and which detected agents to install into.
npx skills add Jaycheng1103/chatgpt-video-editing-skills --full-depthIf you already know you want everything, the repository documents a non-interactive form. The README warns that --all selects both all skills and all detected agents, so confirm that scope first.
npx skills add Jaycheng1103/chatgpt-video-editing-skills --all --full-depthYou can also install one skill at a time, which is the sensible first move if you only want the environment check.
npx skills add Jaycheng1103/chatgpt-video-editing-skills --skill chatgpt-video-editing-setup --full-depthA manual clone followed by a local install is documented for people who want to read the contents first. The README notes that a manual copy must bring each skill's references/ directory along, not just SKILL.md.
git clone https://github.com/Jaycheng1103/chatgpt-video-editing-skills.git
cd chatgpt-video-editing-skills
npx skills add . --full-depthUpdates and removals go through the same tool. If you installed globally, add --global to either command.
npx skills update chatgpt-video-editing-setup chatgpt-short-video-editor
npx skills remove chatgpt-video-editing-setup chatgpt-short-video-editorFor a first run, the README suggests prompting the setup skill to check only and list required changes before installing anything, with no upload or transcription. Once the environment passes, the editing prompt names a full path, a duration range and an aspect ratio, and asks for material inspection and transcription before the strategy. The repository also ships examples/完整提示詞.md for supplying a complete specification in one message.
Where the pipeline stops, and why that is the point
The clearest limitation is the dependency chain. The README lists Git, Python, uv, FFmpeg, ffprobe, the full video-use repository, the Source Han Sans TW OTF fonts, ElevenLabs credentials with available quota, and Pillow. HyperFrames is optional but drags in Node.js 22 or newer and Bun when an approved strategy needs HTML, CSS or GSAP animation. That is a lot of surface area for a tool whose job is cutting video.
When something is missing, the editor skill halts in a safe position, explains the gap and hands off to the setup skill. It will not clone, update, install packages or modify your skills directory on its own. That is a defensible choice and also an annoyance: a missing font turns into a two-skill conversation instead of one command.
The transcription dependency is the other hard edge. The full-precision path assumes ElevenLabs Scribe v2 credentials and quota, and the README requires the agent to state the specific filename, the purpose, and the possible cost before the first upload. If you decline cloud upload, the documented alternative is a local Whisper fallback with lower timestamp confidence that requires extra playback checks at every cut boundary. The README is explicit that this must not be described as equally precise. So the project does not solve offline editing; it offers a degraded mode and tells you it is degraded.
The setup skill also refuses to touch a dirty repository. If an existing checkout has uncommitted changes, it stops rather than pulling, resetting or overwriting, and it will not re-download or overwrite existing font files. Reasonable, and it means a half-finished manual install can block the automated repair you were hoping for.
How it differs from a hosted ChatGPT video editor
The obvious comparison is a hosted editor that takes a prompt and returns a rendered clip, or a general assistant asked to write FFmpeg commands. Those tools own the pipeline. You hand over the file, something comes back, and the intermediate state is invisible.
This repository inverts that. It is a set of instruction documents for an agent you already run, and the artifacts stay on your disk in edit/. There is no service to sign into, no upload of the source beyond the transcription call, and no vendor holding your project state. The trade is that you supply the agent, the machine and the toolchain. A hosted editor needs none of that.
The second difference is the approval structure. A hosted editor renders what it renders. Here, the strategy must be approved in four to eight sentences before cut points exist, and the 720p preview must be approved before a 1080x1920 master is produced. That is slower by design. If your workflow is thirty clips a week with a fixed template, this will feel like friction. If your workflow is one carefully reviewed piece where a bad cut costs you a reshoot, the gates are the feature.
The third difference is language. A generic editor does not know you want Source Han Sans TW burned in for Traditional Chinese subtitles. This one installs that font from the official release branch and composites it at the subtitle stage.
Licensing, third-party terms and what updates cost you
The repository is MIT licensed, and the README states it contains no third-party upstream code. The tools it wires together keep their own terms: FFmpeg, video-use, HyperFrames, ElevenLabs, Pillow and the Adobe Source Han Sans fonts are all governed by their own licences and service agreements, with details in THIRD_PARTY_NOTICE.md. The subtitle font is SIL Open Font License 1.1. None of that is legal advice, and the practical implication is that your ElevenLabs usage is billed by ElevenLabs under its own terms, not by this project.
The project is a community-maintained unofficial effort. The README states plainly that it has no affiliation, authorization, endorsement or partnership with OpenAI, ChatGPT, video-use, HyperFrames, ElevenLabs, FFmpeg, Pillow or Adobe, and that product names and trademarks belong to their respective owners. The last push to the repository was on 2026-07-26, so it is not archived, but there is no release history to speak of and no versioned changelog in the repository.
Upgrade cost is low in one sense: npx skills update refreshes both skills, and --global is the only flag change if you installed globally. The real cost sits upstream. The README says actual requirements follow the current official documentation for video-use, which means an upstream change to that repository can invalidate the environment the setup skill just verified. Budget for re-running the setup skill after any upstream move rather than assuming the check stays green.
Editorial conclusion
Adopt it if you already have a compatible agent that can run shell commands, you are editing Traditional Chinese short video, and you want the source file protected and every stage approved before the next one starts. Do not adopt it if you want a one-click editor, if you refuse to send audio to ElevenLabs and cannot accept lower-confidence timestamps, or if you are not comfortable letting an agent install tools. Before you start, confirm your agent can read and write local files and run terminal commands, check that the full video-use repository is present rather than a single SKILL.md, and decide up front whether you are using Scribe v2 or the local Whisper fallback, because that choice changes how much boundary QA the edit needs.
Frequently asked questions
Is ChatGPT good for video editing?
This project's answer is that the assistant is only as good as the workflow wrapped around it. The repository exists because a model will describe an edit instead of performing and verifying one, so it imposes eight ordered steps, a 720p preview approval gate, and a separate QA pass on the 1080x1920 master before delivery.
What are the best ChatGPT prompts for video editing?
The README does not rank prompts, but it does give two concrete ones. For environment work it suggests asking the setup skill to check only, list required changes, and wait for confirmation before installing. For editing it suggests naming a full source path, a 60 to 90 second target, a 9:16 ratio, and asking for transcription and a strategy before any cutting.
What skills are needed for video editing?
The repository covers the tool side rather than the craft side. It installs and verifies video-use, FFmpeg, ffprobe, the Source Han Sans TW subtitle font, ElevenLabs credentials and optionally HyperFrames, and it assumes you can supply a compatible agent that reads and writes local files and runs terminal commands.
Community notes