watch

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

/watch (Watch Skill)

/watch（Watch Skill）

You don't have a video input; this skill gives you one. It is a thin wrapper around the

watch-skill

CLI — all logic lives in the engine, so this skill works identically on every harness (Claude Code, Codex, Cursor, ...).

This is a drop-in upgrade of the classic claude-video

/watch

skill: same invocation shape, plus a persistent index (

ask

answers follow-ups without re-processing), OCR on frames, scene-aware sampling with perceptual dedup, local Whisper (offline by default, no API key needed), and THE LOOP (capture -> critique -> fix -> re-capture) for iterating on your own output.

你无需提供视频输入；该Skill会为你处理视频输入。它是

watch-skill

CLI的轻量封装——所有逻辑都在引擎中，因此该Skill在所有运行环境（Claude Code、Codex、Cursor等）中的表现完全一致。

这是经典claude-video

/watch

Skill的无缝升级版：调用方式保持不变，新增了持久化索引（

ask

命令可直接回答后续问题，无需重新处理视频）、帧OCR识别、带感知去重的场景感知采样、本地Whisper（默认离线运行，无需API密钥），以及用于迭代优化输出的THE LOOP流程（捕获→评审→修复→重新捕获）。

Step 0 — Preflight (first invocation per session)

步骤0 — 预检（每个会话首次调用）

bash

watch-skill doctor --json

Exit 0 → proceed silently. Do NOT announce that setup is fine.
Non-zero → the JSON lists each failing check with a
```
fix
```
.
```
doctor
```
auto-bootstraps ffmpeg and yt-dlp into a managed bin dir on Windows/macOS/ Linux; re-run once after it reports fixes. Only involve the user when a check still fails after remediation.

watch-skill

itself is not on PATH:

pip install watch-skill

(or

uv tool install watch-skill

), then re-run the doctor.

No API key is required for acquisition, transcription, OCR, indexing, or search: transcription falls back to local faster-whisper. Visual synthesis and verification can use the user's existing Anthropic, OpenAI, Gemini, or OpenRouter key, or an optional local Ollama model. The agent and provider are independent; see the

configuring-vision

skill. Cloud STT is opt-in (

--cloud-stt

) and only ever uploads extracted mono audio — the video file never leaves the machine.

bash

watch-skill doctor --json

退出码0 → 静默继续。无需告知用户配置正常。
非零退出码 → JSON结果会列出每个检查失败项及对应的
```
fix
```
方案。
```
doctor
```
命令会自动在Windows/macOS/Linux系统的管理bin目录中安装ffmpeg和yt-dlp；在它报告修复方案后重新运行一次。仅当修复后检查仍失败时，才需要用户介入。
如果
```
watch-skill
```
未在PATH中：执行
```
pip install watch-skill
```
（或
```
uv tool install watch-skill
```
）安装，然后重新运行doctor命令。

视频获取、转录、OCR识别、索引或搜索均无需API密钥：转录会 fallback 到本地的faster-whisper。视觉合成与验证可使用用户现有的Anthropic、OpenAI、Gemini或OpenRouter密钥，也可选用本地Ollama模型。Agent与服务提供商相互独立；详情请查看

configuring-vision

Skill。云STT为可选功能（需添加

--cloud-stt

参数），且仅会上传提取的单声道音频——视频文件永远不会离开本地设备。

Step 1 — Watch

步骤1 — 处理视频

Parse the user input into source + optional question, then:

bash

watch-skill watch "<source>" [--start T --end T] [--max-frames N] [--transcript-only]

Any yt-dlp-supported site (1800+), direct media URLs, HLS/DASH manifests (
```
--duration 60
```
bounds live streams), and local files all work.
```
--start
```
/
```
--end
```
(
```
SS
```
,
```
MM:SS
```
,
```
HH:MM:SS
```
) switch to dense focused sampling of that window — use for "what happens at 2:30?" questions and for any video over ~10 minutes when the user cares about one section.
```
--timestamps T1,T2,...
```
pins frames at transcript-flagged moments ("look here", "as you can see") that visual selection may miss.
```
--transcript-only
```
skips frames entirely (fastest; no video download when captions exist).
```
--max-frames N
```
tightens the token budget (default: duration-tiered, hard cap 100, max 2 fps).

The report prints an

Indexed: video_id ...

line, frames with

t=MM:SS

timestamps, OCR text, and the transcript.

将用户输入解析为视频源 + 可选问题，然后执行：

bash

watch-skill watch "<source>" [--start T --end T] [--max-frames N] [--transcript-only]

支持所有yt-dlp兼容的网站（超过1800个）、直接媒体URL、HLS/DASH清单文件（
```
--duration 60
```
参数可限制直播流时长），以及本地文件。
```
--start
```
/
```
--end
```
参数（格式为
```
SS
```
、
```
MM:SS
```
、
```
HH:MM:SS
```
）会切换为对指定时间段进行密集聚焦采样——适用于“2:30时发生了什么？”这类问题，以及用户仅关注某一章节的超10分钟长视频。
```
--timestamps T1,T2,...
```
参数会在转录标记的时刻（如“看这里”、“如你所见”）固定提取帧，避免视觉选择遗漏这些关键节点。
```
--transcript-only
```
参数会完全跳过帧提取（速度最快；若视频有字幕则无需下载视频文件）。
```
--max-frames N
```
参数可收紧token预算（默认：按视频时长分级，硬上限为100帧，最大帧率2fps）。

报告中会打印

Indexed: video_id ...

行、带

t=MM:SS

时间戳的帧、OCR文本以及转录内容。

Step 2 — Read the frames

步骤2 — 读取帧

Read every frame path the report lists, in a single message (parallel Read calls), so you see them together in chronological order.

在一条消息中读取报告列出的所有帧路径（并行Read调用），以便按时间顺序查看所有帧。

Step 3 — Answer

步骤3 — 回答问题

Answer from frames + OCR + transcript, citing timestamps. No question → summarize structure, key moments, notable visuals, spoken content.

结合帧、OCR文本和转录内容回答问题，并标注时间戳。若无问题，则总结视频结构、关键节点、显著视觉元素和语音内容。

Follow-ups — use the index, not re-processing

后续问题 — 使用索引，无需重新处理

The video is already indexed. For any follow-up question in this or a LATER session:

bash

watch-skill ask <video_id> "<question>"      # self-healing answer + evidence
watch-skill search "<phrase>"                 # across every video ever watched

ask

(v0.6) answers text-first with timestamped evidence, a confidence score, and a

~N tokens saved

line. It escalates on its own when unsure (dense re-sampling, zoom-crop re-OCR) and states plainly when the video does not clearly show the answer — trust that refusal; do NOT invent an answer past it. Frame paths are listed only when the engine wants you to look yourself (or pass

--frames

); Read them then. Never re-run

watch

for a follow-up on an already-indexed video.

If the user corrects one of your video answers, report it so the system learns (locally):

bash

watch-skill lessons add <video_id> "<question>" "<your wrong answer>" "<the correction>"

视频已建立索引。针对本次或后续会话中的任何后续问题：

bash

watch-skill ask <video_id> "<question>"      # 自修复回答 + 证据
watch-skill search "<phrase>"                 # 检索所有已处理过的视频

ask

命令（v0.6版本）优先以文本形式回答，附带带时间戳的证据、置信度评分，以及

~N tokens saved

行。当不确定答案时，它会自动升级处理（密集重新采样、放大裁剪后重新OCR）；若视频未清晰展示答案，会直接说明——请信任该拒绝结果，切勿编造答案。仅当引擎希望你自行查看时（或添加

--frames

参数）才会列出帧路径；此时再读取这些帧。对于已建立索引的视频，后续问题绝不要重新运行

watch

命令。

如果用户纠正了你对视频的回答，请报告该信息以便系统本地学习：

bash

watch-skill lessons add <video_id> "<question>" "<your wrong answer>" "<the correction>"

THE LOOP — iterate on your own output

THE LOOP — 迭代优化自身输出

When the user asks you to fix UI/visual output and verify the fix:

bash

watch-skill loop start "<url-or-screen:-or-file>" "<pass criteria>" [--script '<json steps>']

当用户要求你修复UI/视觉输出并验证修复效果时：

bash

watch-skill loop start "<url-or-screen:-or-file>" "<pass criteria>" [--script '<json steps>']

... you apply the suggested fixes ...

... 你应用建议的修复方案 ...

watch-skill loop iterate <loop_id>


The critique returns structured issues with timestamps and suggested fixes.
YOU change the code; the loop only observes. On pass it renders a
before/after MP4+GIF proof. `watch-skill capture "<target>"` records without
critiquing.

watch-skill loop iterate <loop_id>


评审结果会返回带时间戳的结构化问题及建议修复方案。由你修改代码；loop仅负责观察。验证通过后，会生成前后对比的MP4+GIF证明文件。`watch-skill capture "<target>"`命令仅录制内容，不进行评审。

Security posture

安全态势

The video file itself NEVER leaves the machine. Only extracted mono-16kHz audio may go to a cloud STT API, and only with explicit
```
--cloud-stt
```
.
No cookies, no logins — only public data is requested.
API keys live in env vars /
```
.env
```
; they are never logged or echoed.
Downloads are cached under
```
~/.watch-skill/cache
```
(LRU, size-capped).

视频文件本身永远不会离开本地设备。仅当明确添加
```
--cloud-stt
```
参数时，提取的16kHz单声道音频才可能上传至云STT API。
无需cookie或登录——仅请求公开数据。
API密钥存储在环境变量或
```
.env
```
文件中；永远不会被记录或回显。
下载内容会缓存到
```
~/.watch-skill/cache
```
目录（采用LRU算法，有大小限制）。