video-transcript

name

video-transcript

description

Use this skill whenever the user provides a YouTube URL and wants the spoken content as a transcript or Word document. Triggers include: any YouTube link combined with words like '講稿', '字幕', '逐字稿', '整理', 'transcript', 'script', 'transcribe', '幫我整理', '轉成文字', or '翻譯'. Also trigger when the user pastes a youtu.be / youtube.com URL alongside any request to read, summarize, or document the content — even if they don't use exact keywords. Additionally trigger when the user pastes a block of subtitle/caption text (with or without timestamps) and asks to turn it into a clean document or translate it. Always use this skill for any YouTube-to-document workflow — do NOT handle it ad hoc.

Video Transcript Skill

Extracts spoken content from YouTube videos and delivers polished Word (.docx) documents:

• 逐字稿 — verbatim, timestamped
• 整理版 — cleaned up, sectioned, with headings
• For non-Chinese videos: both an 原文版 and a 正體中文版 (two separate .docx files)

---

Step 1 — 取得字幕文字

Claude.ai 環境說明

Claude.ai 網頁版無法直接抓取 YouTube 字幕 API（fetch 受環境限制）。正確流程是由使用者複製字幕文字後貼入對話。

判斷使用者是否已貼上文字

情況 A：使用者只貼了 YouTube URL，沒有附上文字

立即回覆以下引導說明（語氣親切、步驟清楚）：

收到影片連結！要製作講稿，需要先把字幕文字複製給我。以下是最快的方式：

方法一：YouTube 內建字幕複製（最簡單，30 秒完成）

1. 在電腦瀏覽器開啟影片
2. 點影片下方「⋯」（更多）→「開啟逐字稿」
3. 右上角點「⋮」→「切換時間戳記」（可關閉時間戳，讓文字更乾淨）
4. 全選文字（Ctrl+A）→ 複製（Ctrl+C）→ 貼到這裡

方法二：沒有「開啟逐字稿」選項？ 表示影片沒有字幕。可改用：

• tactiq.io（免費，開啟影片後自動產生字幕，可直接複製）
• downsub.com（貼入 YouTube URL 即可下載字幕文字檔）

把複製好的文字貼回來，我馬上幫你製作 Word 講稿 ✦

Stop here and wait. Do not proceed without actual transcript text.

---

情況 B：使用者同時提供了 URL 和文字內容

直接進入 Step 2，不需要再詢問。

---

情況 C：使用者貼了字幕文字但沒有 URL

也直接進入 Step 2。記錄影片標題為「未知」，在文件 header 說明來源不明。

---

Step 2 — Detect Language & Plan Output

Once you have the raw transcript text:

1. Detect the primary language of the content (not the metadata/title — the spoken content).
2. Decide output plan:

語言	輸出文件
繁體/簡體中文	2 個 .docx：逐字稿、整理版（均為中文）
英文 / 日文 / 其他	4 個 .docx：原文逐字稿、原文整理版、正體中文逐字稿、正體中文整理版

Tell the user the plan before proceeding:

偵測到影片語言為【English】，將產出 4 份文件：原文逐字稿、原文整理版、正體中文逐字稿、正體中文整理版。

---

Step 3 — Prepare Both Text Versions

逐字稿版 (Verbatim)

• Keep all words, including filler words (um, uh, 那個, 就是) — but remove repeated stutters like "I I I"
• Add timestamps every ~2 minutes: [00:02:15]
• Do NOT add headings or restructure
• Minimal cleanup only: fix obvious ASR errors where context is unambiguous

整理版 (Edited)

• Remove filler words, false starts, repetitions
• Break into logical sections with Heading 2 titles (infer from content)
• Write a brief 摘要 (Summary) at the top (3–5 sentences)
• Fix punctuation and sentence breaks
• Preserve the speaker's meaning and terminology faithfully — do not paraphrase aggressively

正體中文版 (if needed)

• Translate the above two versions into 正體中文 (Traditional Chinese)
• Use natural, fluent 繁體中文 — not word-for-word translation
• Preserve technical terms; add original term in parentheses on first use: 機器學習 (machine learning)
• Timestamps in 逐字稿 should be preserved

---

Step 4 — Create .docx Files

Use the docx npm package. For each document, follow the docx skill conventions.

Document Structure

Header (all documents):

影片來源：[YouTube URL]
影片標題：[title from page metadata if available]
產出日期：[today's date]
文件類型：逐字稿 / 整理版
語言：原文 (English) / 正體中文版

Page setup: A4, margins 2cm all sides (2835 DXA), font: "Microsoft JhengHei" for Chinese content, "Arial" for English content.

Styles:

• Title: 18pt bold
• Heading 1: 14pt bold (document sections like 摘要, 逐字稿內容)
• Heading 2: 12pt bold (content sections in 整理版)
• Body: 11pt, line spacing 1.5
• Timestamps: 9pt, grey color 999999, inline with text

File naming:

[VideoTitle]_逐字稿.docx
[VideoTitle]_整理版.docx
[VideoTitle]_EN_逐字稿.docx       ← only for non-Chinese videos
[VideoTitle]_EN_整理版.docx       ← only for non-Chinese videos

Use the video title (up to 30 chars, sanitized — no /\:*?"<>|).

docx-js scaffold (A4, Traditional Chinese)

const { Document, Packer, Paragraph, TextRun, HeadingLevel, AlignmentType } = require('docx');
const fs = require('fs');

const doc = new Document({
  styles: {
    default: { document: { run: { font: "Microsoft JhengHei", size: 22 } } },
    paragraphStyles: [
      { id: "Heading1", name: "Heading 1", basedOn: "Normal", next: "Normal",
        run: { size: 28, bold: true, font: "Microsoft JhengHei" },
        paragraph: { spacing: { before: 240, after: 120 }, outlineLevel: 0 } },
      { id: "Heading2", name: "Heading 2", basedOn: "Normal", next: "Normal",
        run: { size: 24, bold: true, font: "Microsoft JhengHei" },
        paragraph: { spacing: { before: 180, after: 90 }, outlineLevel: 1 } },
    ]
  },
  sections: [{
    properties: {
      page: {
        size: { width: 11906, height: 16838 },
        margin: { top: 1134, right: 1134, bottom: 1134, left: 1134 }
      }
    },
    children: [ /* paragraphs here */ ]
  }]
});

Packer.toBuffer(doc).then(buf => fs.writeFileSync("output.docx", buf));

Timestamp rendering (逐字稿)

// Inline timestamp before a paragraph
new Paragraph({
  children: [
    new TextRun({ text: "[00:02:15] ", color: "999999", size: 18 }),
    new TextRun({ text: "實際講稿內容從這裡開始...", size: 22 }),
  ],
  spacing: { line: 360 }  // 1.5 line spacing
})

Validate each file

python /mnt/skills/public/docx/scripts/office/validate.py output.docx

---

Step 5 — Deliver to User

1. Copy all output files to /mnt/user-data/outputs/
2. Call present_files with all files
3. Give the user a brief summary:
   • Video title and detected language
   • Number of files produced
   • Approximate word count of the transcript
   • Any caveats (e.g., "部分時間戳可能有誤差，建議對照原影片確認")

---

Edge Cases

情況	處理方式
字幕為自動產生（ASR）	在文件 header 標註「字幕來源：YouTube 自動產生，可能含辨識錯誤」
影片超過 60 分鐘	整理版每 15 分鐘為一個 Heading 1 大節
多語言混合（如台灣中英夾雜）	以主要語言判斷；在正體中文版保留英文專有名詞
字幕有廣告/贊助段落	在逐字稿保留，整理版可標註「[贊助段落]」並縮排
影片為 Shorts（< 3 分鐘）	只產出整理版一份，不做逐字稿（告知用戶）