Video Watch Plus

看视频与听声音：把视频/GIF/录屏同时翻译成「可核算的图版序列」与「可信度经过检验的音轨结论」，产出带时间戳的观看交付。视觉侧 ffmpeg 逐帧量测（切点/冻结/运动/像素突变/静音）→ 骨架优先采样 → 图版与位置映射 → 画面文字裁切放大读取；听觉侧先做声学侦察判定有无语音，再做 ASR 转写并做幻觉证伪；最后三路证据互校。含 Windows 一键环境安装与上游 vw.py 的两处 Windows 缺陷修补。Invoke when 用户要求看视频/看动图/分析录屏/GIF 拆帧/视频里发生了什么/某时刻画面是什么/把视频里的文字或语音转写出来/这条音轨里有没有人说话。

17 starsImages, Video & Audio

Install

skills CLI (Claude Code, Cursor, Codex and more)

npx skills add xinhailaopo/video-watch-plus --skill video-watch-plus

Manual install

git clone --depth 1 https://github.com/xinhailaopo/video-watch-plus.git /tmp/video-watch-plus
mkdir -p ~/.claude/skills
cp -r /tmp/video-watch-plus/. ~/.claude/skills/video-watch-plus

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# video-watch-plus — 看视频与听声音

把连续时间轴同时翻译成两份**可核算**的东西：一份是**离散图版序列**（给只有单帧视觉能力的模型），
一份是**经过检验的音轨结论**（语音 / 音乐 / 静音，以及可信的转写）。

**核心机理**：ffmpeg 看全部帧（零 token），我看其中几十帧，二者靠**时间戳**缝合；
声学层用纯物理量先判"有没有人声"，再让 ASR 回答"说了什么"。

**交付目标**：给出**人能用的观看结论**（完整描述 + 细节 + 带时间节点的转写），不是抖动审计。

本文件写给**执行这个技能的智能体**：下文"我"即执行者。

---

## 两条铁律（违反其一，结论就会是编的）

### 铁律一：图版里的文字一律不采信

实测两次：3×3 图版里 16px 的时间码我"读"出了九个不同值（回原图裁切放大后发现九格字面全同）；
另一次我"读"出一组结构规整的标题，放大后发现画面上根本没有这些字。

> **视觉在分辨率不足时不报"读不到"，它渲染一个合理内容。**

### 铁律二：音频域同样会幻觉，而且形态更隐蔽

实测：对一段**纯 BGM** 视频关掉 VAD 强制出字，Whisper 对三个**互不重叠**的时间窗输出了
**完全相同**的伪造字幕，而模型自报的 `no_speech` 概率高达 0.44–0.74。

> **没有语音时，ASR 会拿最常见的模板句填空。转写文本非空 ≠ 有语音。**

判定规则：**只要发现互不重叠的时间窗输出了相同文本，无论多通顺，一律判为幻觉。**

---

## 环境

```powershell
# Windows 一键装（ffmpeg 完整构建 + ImageMagick 7 便携版 + 用户级配置 + 环境变量）
powershell -ExecutionPolicy Bypass -File scripts\setup-windows.ps1 -VwPy vendor\vw.py
python vendor\vw.py doctor
```

macOS / Linux：`brew install ffmpeg imagemagick` 或 `apt install ffmpeg imagemagick`，
再把 `vendor/vw.py`、`scripts/audio_probe.py` 的 ffmpeg 路径指过去即可。

音频侧可选依赖：

```bash
pip install faster-whisper "av<19" numpy
```

> **`av<19` 不能省。** faster-whisper 1.2.1 调用 `av.open(..., metadata_errors=...)`，
> 而 PyAV 19 已移除该参数，装最新的 av 会在转写时直接 `TypeError`。
> 国内环境还需 `HF_ENDPOINT=https://hf-mirror.com` 与 `HF_HUB_DISABLE_XET=1`
> （后者是因为 huggingface_hub 默认走 Xet 传输，其域名不可达会报 401）。setup 脚本已写好这两个变量。

---

## 工作流

### 视觉侧

Read the full skill on GitHub

Details

Skill name: video-watch-plus
Repository: xinhailaopo/video-watch-plus
Repo updated: 4 October 2026

FAQ

What does the Video Watch Plus skill do?

How do I install the Video Watch Plus skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add xinhailaopo/video-watch-plus --skill video-watch-plus. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Video Watch Plus popular?

It lives in xinhailaopo/video-watch-plus, which has 17 GitHub stars and 0 forks.

Related skills

Roblox Game Builder

YGMontop/AutoGameSkill

Builds a complete Roblox game from a one-line idea. Interviews the user with many questions, writes a design doc for approval, has the user sign in through the integrated browser, then makes the images, gamepasses and developer products, 3D models, uploads, map and code, tests it, and publishes. Use whenever the user gives a Roblox game idea or says things like "make me a Roblox game", "build a tycoon/obby/simulator/horror game", or "turn this idea into a full game".

Images, Video & Audio

Libro 3d Interactivo

csepulvedai/libro-3d-interactivo

Construye libros educativos interactivos en 3D tipo pop-up (estilo "Every Joule"): un libro de tapa de tela sobre un escritorio, con páginas HTML reales y, en cada pliego, un diagrama a tinta que se levanta de la página para que el estudiante Prediga → Pruebe → resuelva un Desafío y gane un sello. Hecho con Three.js, se entrega como un único index.html (sin internet, listo para hosting o Moodle) más notas docentes, con narración por pliego en voz chilena (Gemini TTS, verificada transcribiendo) y un cursor guía que hace el gesto justo cuando la voz nombra cada botón. Sirve para cualquier asigna

Images, Video & Audio

Video Script Builder

erduo1998-cell/video-script-builder

Convert a Chinese video voiceover script (逐字稿、口播稿) into a precise, renderer-bound storyboard specification for HyperFrames or Remotion, or revise an existing video-spec-hf.md or video-spec-remotion.md. Use when the user asks for 分镜、画面设计、镜头拆解、HyperFrames/Remotion 实现规划、动画时间轴、字幕或转场规划. Produce exactly one selected storyboard spec; never render video or create a composition.

Images, Video & Audio

Movie Maker Fast

AEON-7/aeon-movie-maker

Fast cinematic video generator built on LTX 2.3 (distilled fp8, video-only) + abliterated Gemma text encoder + physics/control/motion LoRA stack. Produces a full screenplay-driven film at roughly 10-15x the speed of the WAN MultiTalk pipeline (Movie Maker Slow WAN). Use this skill when the user wants a movie / film / cinematic video / animated scene production and speed matters more than strict lip-sync accuracy. Companion audio pipelines (Qwen3-TTS Three-Lock for dialogue, ACE Step XL base via music_maker for music, MMAudio for SFX) compose alongside via the stitcher's audio mux. Distinct fro

Images, Video & Audio

Google Flow Scripting

agbelemi/google-flow-scripting-skill

Plans, scripts, storyboards, validates, and audits multi-shot Google Flow, Veo, and Gemini Omni Flash video projects. Use for narrative shorts, animation, live action, documentaries, ads, music videos, continuity planning, reference-image planning, prompt writing, or pre-generation QA.

Images, Video & Audio

Construction Takeoff

hamzaabduljabbar/claude-construction-takeoff-skill

Run a quantity takeoff on a construction drawing set that has already been processed by the Drawing Analyzer skill. Identifies every countable or measurable element, marks them up on the drawings for audit, calculates primary and secondary quantities (excavation volume, concrete volume, formwork area, reinforcement weight), and exports the results to an Excel workbook with confidence levels. Trigger on phrases like "run a takeoff", "quantity takeoff", "use my construction takeoff skill", or any request to extract quantities from a drawing set.

Images, Video & Audio

Browse the most popular Claude skills, all images, video & audio skills, or write your own with the skills generator.