SeeCut skill
A Claude Code skill that turns a talking-head video into a short motion-graphics edit and checks its own cut
- 378 ★ on the repository
- Last commit
Extended access
Why this level
- The skill runs scripts, renders video and installs dependencies on your computer
- Video and frames for checking go to Gemini through the Antigravity CLI, and you sign in to a Google account
- The agent takes web page screenshots through playwright-core
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/seecutDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The skill runs scripts, renders video and installs dependencies on your computer
- Video and frames for checking go to Gemini through the Antigravity CLI, and you sign in to a Google account
- The agent takes web page screenshots through playwright-core
This is third-party code. Review the repository files before installing.
What it does
SeeCut builds a short video from a vertical clip of a person or digital avatar talking and screenshots of the things they mention. Cards, callouts and hand-drawn marks appear in the frame, and the face switches between full screen and a small window. The agent studies the material, finds real evidence for the frames and renders the scene in HyperFrames. The rules forbid fake interfaces to fill a frame. The central idea is that the model watches its own result. SeeCut connects Gemini through the Antigravity CLI, and it watches the whole video without splitting it into frames. Each new version is compared with the previous one, and the better one stays. The output is a render, a layered project package and a report on what changed. With the Jianying draft engine installed, the agent builds a draft that opens in the editor as separate tracks. The repository also carries scripts for an optional first step: Doubao TTS voice-over and a HeyGen AV5 avatar.
Who it is for. For short-video creators and marketers who film a talking head and want a motion-graphics edit without manual editor work.
Good fit when
- You have 20–30 seconds of vertical half-length footage and 3–6 screenshots of what the speech mentions
- You want a version of the video where cards and callouts follow each phrase
- You want to hand the result to a video editor for fine edits track by track
- There is no footage of a person and you need a digital avatar: the digital-human folder has scripts and step notes
- You want to see a sample on the first seconds of a long video before running the full one
Not a fit when
- You need only subtitles, only voice-over or only generated video: the skill is not built for that
- You have no access to the Antigravity CLI and a Google account: without it the self-check and iteration will not run
- You need a result in minutes: one 25-second sample takes about an hour
- You work on Windows or Linux: it was tested only on macOS
Example request
/seecut material: ~/Videos/launch-talk (a 25-second vertical recording and four screenshots)Limitations
Tested on macOS only. Of the agents Claude Code is tested; Codex and WorkBuddy should work in theory. You need the Antigravity CLI (the agy command) signed in to a Google account, the HyperFrames CLI (npx -y hyperframes@latest), ffmpeg, Node 22 or newer, Python 3, the faster-whisper package and playwright-core in the project folder. The preflight.sh script checks the environment.
A free Google account has a weekly quota that may not cover a full iteration loop. The Google account region must be a supported one. One loop produces 2–3 versions, each rendered and compared with the last, so a 25-second clip takes about an hour and longer ones more. For material over 60 seconds the skill asks you to try the first 20–30 seconds first. The README and the skill are in Chinese, with English and Japanese translations.
How to disable. Delete the seecut folder from ~/.claude/skills. Finished videos sit in Downloads, and the layered packages and reports live in the material folder you named.
FAQ
How do I run SeeCut in Claude Code?
Copy the skill/seecut folder into ~/.claude/skills, run zsh skill/seecut/scripts/preflight.sh and wait for the required items to turn green. Then, in a new conversation, type /seecut and the path to your material folder. You get a render in Downloads, a layered package and a report of the versions.
Why does it need the Antigravity CLI?
Through it Gemini watches the whole video like a viewer, while Claude and GPT usually see a handful of frames and miss pace and animation. Without the Antigravity CLI there is no self-check and no iteration loop. Headless calls need the --dangerously-skip-permissions flag.
How does SeeCut decide that a version got better?
A version first passes a hard-fault check, then is compared pairwise with the previous one. The author found that absolute model scores are unreliable while picking the better of two works, so a version is accepted when it wins in both display orders. At most three versions, and each must visibly improve.
Can I open the result in an editor?
Yes, as an optional third step. The skill outputs a layered project package, and with the Jianying draft engine installed it builds a draft with separate tracks where sound effects sit on their own track. Without the engine the effects are mixed into the final video.
What is the digital avatar in the digital-human folder?
An optional first step: text is voiced with Doubao TTS, then HeyGen AV5 makes a vertical half-length clip. The author rejected HeyGen's built-in voices because they sound artificial. The scripts doubao_tts.py and heygen_avatar.py sit in the folder, and the tools can be swapped.
Discussion
No questions or reviews yet.
Related
SkillsAgentic video studio for Claude Code, Cursor and Codex: script, voiceover, images, stock footage, music, subtitles and render
Skills and CLI from HeyGen: the agent writes HTML compositions and renders them into MP4 videos, captions, motion graphics and decks
GPT-Image2 Style Library
gpt-image-2-style-library
A skill picks a matching template and style from a library of GPT-Image2 cases and assembles it into a finished prompt
Blender MCP
MCP for Blender
An MCP server and Blender addon: the agent creates and edits objects, materials and scenes, pulls assets and runs Python in Blender
A video editing skill for Claude Code and Codex: the agent cuts pauses and filler words, grades color, burns subtitles and adds animated overlays
An HTML design skill: clickable prototypes, slides with PPTX export, MP4 and GIF animations, infographics and expert review
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| SeeCut (this entry) | no | 378 | |
| video-use | no | 29,616 | cuts pauses and filler words from recorded video and adds captions; SeeCut builds motion graphics around a talking head |
| AutoClip | no | 9,663 | cuts long videos from a link or file into short clips, with no motion graphics or self-check |
README in short
Two videos come first for comparison: the original avatar and the edited result. The quick start uses a sample material folder, and a three-step pipeline runs from the avatar to a Jianying draft.
README badge
Original SKILL.md
--- name: seecut description: 网感口播精剪（v3.4）：口播视频加动效，自动剪成"网感口播精剪风格"竖版/横版样片，交付渲染成片+分层工程包（装了剪映引擎则再出剪映分层草稿，音效放草稿；没装则音效混进成片）。当用户给了口播粗剪(+B-roll/截图/录屏)并想要"口播动效/加卡片弹字/剪成短视频/做动效样片/网感口播精剪"时用。自循环:预检→素材理解→找真证据→编排→HyperFrames构建→硬门+两两对比自迭代(≤3版)→配音效→交付→用户终审。不做:纯字幕/纯配音/纯生图/BGM。 --- # SeeCut v3.4 · 网感口播精剪（口播动效自动剪辑） 把口播粗剪 + 素材，剪成"网感口播精剪风格"动效样片：数字人/真人口播打底，把画面让给高密度信息层（**真实证据** + chip/卡/大字 + 手作质感 + 弹簧动效），人像按"对你说露、给你看藏"调度。**你（执行窗）自己跑完整循环，交两份东西给用户终审：渲染成片（原声，不混音效）+ 剪映分层草稿（带音效轨）。长任务每 15 分钟往 runlog 写一行进度。等后台任务一律用 `zsh scripts/wait_for.sh <日志> <ASCII完成标记> <超时秒>`，禁止手写 until/while 循环（Round2-4.4 手写循环空转 10 小时）。** ## 0 · 开工前（必做） 1. **读** `references/00-总纲.md`（16 条，全 skill 最高规则；其它文件与它冲突时以它为准并在报告里指出）+ `references/08-避坑SOP.md`（渲染铁律）。 2. **预检**：`export JIANJI_PROXY=http://127.0.0.1:1087` 后 `zsh scripts/preflight.sh` —— 必需项缺失就先装或报告用户，别跑到一半才发现。hyperframes CLI 走 npx，是必需项（官方组件库靠它）。 3. 建工程目录（素材、`build/`、`qc/`、文档都放这里），`cd build && npm i playwright-core`（别 sudo）。 4. **先样片后全片**：素材超过 60 秒时，先截开头 20-30 秒跑完整循环出样片，交用户确认风格和方向后再跑全片（一次循环会出 2-3 版，全片耗时随时长和版本数一起涨，方向错了代价很大）。用户明确说直接出全片则跳过。 ## 1 · 素材理解 → `素材台账.md` - **视频素材**用 agy 读：`zsh scripts/agy_read.sh <文件> "<问题>"`（已含新项目隔离、超时、禁 OCR 子任务；长视频先切 ≤25s 段、降 540p），别只抽帧。**图片素材**执行窗自己读图（更准，也顺便逐张核敏感字、记 fp-manifest）。 - 判人像素材类型（`02` §6.5）：①半身口播（真人或数字人完整半身 A-roll）/ ②角落圆 PIP / ③点缀型数字人。 - 若 `references/金锚集/` 里有你自备的参考片，按功能挑 1-2 条看一遍；没有就跳过。 ## 2 · 转写 → 词级时间轴 faster_whisper(small) 词级转写 → `seg_words.json`，转一次就沉淀。动效锚关键词被说出的那一帧（非句首）。 ## 3 · 找真证据（v2 新增，参考片密度的来源） - 逐句标"有据/无据"：提到真实产品/网页/文档/数据/操作 = 有据。 - 有据句按顺序拿证据：用户素材 → `node scripts/grab_evidence.mjs <素材目录> <url>...` 截真网页（在 build/ 下或工程根目录跑都行）→ 自己录屏；3D 比喻/场景图可 Grok 生成（不能当"真界面"用）。 - **每张截图入片前读图核敏感字**（总纲第 7 条），同时把图里看得清的字逐行记进 `build/fp-manifest.txt`（评委指纹核要用）；拿不到的列进"证据缺口清单"（`04` 第3步），按 `02` §3 第4条退回无据打法。 - **禁假证据**：不许用占位窗口/伪代码网格/假 Mac/假 App 界面冒充或"填满"（总纲第 5 条）。 ## 4 · SHOTBOOK 编排（`references/04-SHOTBOOK施工规格.md`）