yueying · diff
git:20260909.b3458aa to git:20260912.ac8cce6
57 added, 20 removed. Audit A to A.
---
name: yueying
- description: 让 AI 看视频。用户给出本地视频文件路径或视频链接（B站、YouTube、抖音、小红书等），想要总结、提取信息、答疑、照着教程操作、整理笔记时使用。先用 yueying 命令把视频转成文字稿和关键帧，再读结果完成用户的需求。
+ description: Let AI watch videos. Use when the user gives a local video/audio file path or a video URL (YouTube, Bilibili, Douyin, Xiaohongshu, TikTok, Vimeo …) and wants a summary, notes, answers, extracted steps/commands/code, or to follow a tutorial. Prefer the yueying MCP tools when available, otherwise run the yueying CLI to get a timestamped transcript plus keyframes, then read the results. 让 AI 看视频。用户给出本地视频文件路径或视频链接（B站、YouTube、抖音、小红书等），想要总结、提取信息、答疑、照着教程操作、整理笔记时使用。先用 yueying 命令把视频转成文字稿和关键帧，再读结果完成用户的需求。
---
- # 阅影 · 让 AI 看视频
+ # yueying (阅影) · Let AI watch videos
- Claude 读不了视频文件，但能读文字和图片。`yueying` 命令把视频拆成：带时间戳的文字稿（平台字幕优先，没有就本地语音识别，支持中英日韩法德俄等多语言）+ 关键帧图片（场景切换处抽帧，左下角烧了编号和时间）+ 九宫格总览图 + 一份 report.md 索引。
+ You cannot open a video file, but you can read text and images. yueying turns a video into:
+ a timestamped transcript (platform subtitles first; otherwise local faster-whisper speech
+ recognition in dozens of languages), keyframes taken at scene changes (each labelled
+ "#number mm:ss" bottom-left), 3x3 contact sheets, and a `report.md` index. Everything runs offline.
- ## 步骤
+ ## Step 0 — use the MCP tools when they exist
- 1. 运行（第一次跑语音识别会下载模型，可能要几分钟；长视频识别需要一些时间，耐心等）：
+ If MCP tools named `watch_video` / `get_transcript` / `search_transcript` / `get_frames` /
+ `get_frame_at` / `list_videos` from a server called `yueying` are available, use them and skip the
+ CLI steps below:
+ 1. `watch_video(video=<absolute path or URL>)`. If the reply starts with `RUNNING`, call
+ `watch_video` again with the same `video` to keep waiting (it resumes the same job); do not start
+ another video meanwhile. `DONE` contains the overview and the transcript.
+ 2. `get_frames(video=<video_id>)` shows the contact sheets — read these before single frames.
+ `get_frame_at(video, time="03:15")` for one exact moment (code, slides, UI).
+ `get_transcript` for a time range, `search_transcript` to find where something is said.
+
+ ## CLI steps (no MCP server)
+
+ 1. Run (the first speech recognition downloads a model, which can take minutes; long videos take a
+ while — be patient):
+
```
- yueying "<视频路径或链接>" --out "<输出目录>"
+ yueying "<video path or URL>" --out "<absolute output dir>"
```
- - 输出目录建议放在当前项目下，如 `./yueying_out/<视频名>`；不传 `--out` 也会用这个默认值。
- - 视频里没人说话、只想看画面：加 `--no-asr`。只要文字不要画面：加 `--no-frames`。
- - 已知语言可加 `--lang zh` / `--lang en` 等，识别更稳。
- - 画面细节多（代码、PPT、操作演示）想看得更密：加 `--interval 2`（每 2 秒抽一帧）；默认按时长自动，1 分钟内每 2 秒、10 分钟内每 6 秒、更长每 12 到 20 秒。
- - B站 高清或会员视频需要登录：加 `--cookies-from-browser chrome`（或 edge）。
- - 机器慢或没显卡：加 `--model small` 换小模型。
- - 多个视频可以一次传多个路径或链接，各出各的文件夹，`--out` 指定的是父目录，里面会多一份 `index.md` 总目录；B站 分 P / 合集加 `--all` 全部处理。
+ - Use an absolute `--out`, e.g. `<project>/yueying_out/<video-name>`; without `--out` the
+ default is `./yueying_out/<video-name>` relative to the current directory.
+ - Add `--ui-lang en` when the user writes in English (report.md headings in English; the
+ default is Chinese).
+ - Nobody speaks / only the pictures matter: add `--no-asr`. Text only, no pictures: `--no-frames`.
+ - Known language: `--lang zh` / `--lang en` / `--lang ja` … improves accuracy.
+ - Dense visual detail (code, slides, demos): `--interval 2` (one frame every 2 s). The default is
+ automatic by duration: 2 s under 1 min, 6 s under 10 min, 12–20 s for longer videos.
+ - HD or member-only Bilibili, sign-in-gated YouTube: `--cookies-from-browser chrome` (or `edge`;
+ close Chrome first on Windows).
+ - Slow machine or no NVIDIA GPU: `--model small` (or `--model auto`, which picks
+ large-v3-turbo on a GPU and small on CPU).
+ - Several videos at once: pass several paths/URLs; each gets its own folder under `--out` plus an
+ `index.md`. Bilibili multi-part / collections / playlists: add `--all` to process every entry.
- 2. 用 Read 读输出目录里的 `report.md`。里面有简介、章节、画面总览图清单、关键帧清单、按段落带时间戳的文字稿。
+ 2. Read `report.md` in the output folder: summary, chapters, list of contact sheets, list of
+ keyframes, and the transcript in timestamped paragraphs.
- 3. 用 Read 看 `grid_01.jpg` 等总览图，快速了解画面走向。需要看清某一刻的细节（代码、PPT、界面操作），再打开 `frames/` 下对应的单帧大图。文字稿和画面靠时间戳对照。
+ 3. Read `grid_01.jpg` etc. (contact sheets) to get the visual storyline. Open single frames in
+ `frames/` only when you need to read code, slides or UI text at one moment. Match pictures to the
+ transcript by timestamp.
- 4. 按用户的需求提纯：总结、列步骤、抄代码、回答问题、写笔记。引用视频内容时带上时间点，如「（03:15）」，方便用户回看。
+ 4. Deliver what the user asked for: summary, steps, code, answers, notes. Quote the video with
+ timestamps like (03:15) so the user can jump there.
- ## 注意
+ ## Notes
- - 语音识别可能把专有名词、代码听错，画面上出现的文字以画面为准。
- - 文字稿很长时不必全读，先看章节和画面总览，再按需读对应时间段。
- - 如果命令不存在：`pip install yueying` 然后 `yueying --install-skill`；有 NVIDIA 显卡再加 `pip install yueying[cuda]`。
+ - Speech recognition mis-hears names, numbers and code; trust on-screen text from frames over
+ the transcript.
+ - For long transcripts, do not read everything: start with chapters and contact sheets, then read
+ the relevant time range.
+ - If the command is missing: `pip install yueying` then `yueying --install-skill`; on a machine
+ with an NVIDIA GPU also `pip install "yueying[cuda]"`. With uv: `uvx yueying "<path or URL>"`.
+
+ ## 中文提示
+
+ - 有 `yueying` MCP 工具时优先用 `watch_video`；回复以 `RUNNING` 开头就用同一个 `video` 再调一次，不要同时开第二个视频。
+ - 没有 MCP 时运行 `yueying "<视频路径或链接>" --out "<绝对路径输出目录>"`，然后读 `report.md`、`grid_01.jpg` 等总览图，需要细节再看 `frames/` 单帧。
+ - 语音识别可能把专有名词、代码听错，画面上的文字以画面为准；引用视频内容时带上时间点，如「（03:15）」。