Video Transcript Skill

Extract subtitles from a video URL and deliver clean .docx transcript files. For non-Chinese videos, produce two files: original language + Chinese translation. For Chinese videos, produce one file.

Workflow

Step 0 — First-time setup (once only)

Check if node_modules exists in the scripts directory. If not, run:

$skillScripts = "$env:USERPROFILE\.agents\skills\video-transcript\scripts"
if (-not (Test-Path "$skillScripts\node_modules\docx")) {
    Push-Location $skillScripts
    npm install
    Pop-Location
}

Step 1 — Download subtitles with yt-dlp

Run in a temp directory (use $env:TEMP\vt_<random> on Windows):

$tmp = "$env:TEMP\vt_$(Get-Random)"
New-Item -ItemType Directory -Force $tmp | Out-Null

YouTube:

# Try manual subs first, fall back to auto-generated
yt-dlp --skip-download --write-subs --write-auto-subs --sub-langs "en,zh-Hans,zh-Hant,zh" --convert-subs srt -o "$tmp\sub" "<URL>"

Bilibili:

yt-dlp --skip-download --write-subs --sub-langs "zh-Hans,zh,ai-zh" --convert-subs srt -o "$tmp\sub" "<URL>"

Other platforms: Use the same YouTube command — yt-dlp handles most platforms automatically.

After download, list what was fetched:

Get-ChildItem $tmp -Filter "*.srt"

Step 2 — Detect available subtitle files

Pick the best available file:

• Prefer manual subs over auto-generated (.en.srt over .en-orig.srt or .en-auto.srt)
• For Bilibili: pick zh-Hans or ai-zh
• If multiple languages exist, note them all

If NO subtitle file was downloaded:

• Tell the user: "该视频没有可用字幕，yt-dlp 无法提取文字内容。如需转录请使用 Whisper 语音识别（需额外安装）。"
• Stop here.

Step 3 — Parse SRT to plain text

Read the .srt file and strip all timing/index lines. Keep only the spoken text lines.

SRT format to strip:

1
00:00:01,000 --> 00:00:03,500
This is the spoken text.

2
00:00:04,000 --> 00:00:06,000
More text here.

Output: plain text, one paragraph per subtitle block, blank lines between blocks removed (merge into flowing paragraphs). Also strip HTML tags like <i>, <b>, <font ...>.

Do this parsing yourself by reading the file content — no external tool needed.

Step 4 — Translate if needed

Detect the subtitle language:

• If Chinese (zh, zh-Hans, zh-Hant): skip translation, one file only
• If other language: translate the full plain text to Chinese using your own translation capability

Translation guidelines:

• Produce natural, readable Chinese — not word-for-word
• Preserve paragraph structure
• Keep proper nouns, technical terms transliterated or explained inline if needed
• Do NOT add commentary or summaries — pure translation only

Step 5 — Generate .docx files

Use the bundled script. The script path is: ~/.agents/skills/video-transcript/scripts/make_docx.js

Get the video title from yt-dlp output or use a sanitized version of the URL as fallback.

For non-Chinese video (two files):

# Original
node "~/.agents/skills/video-transcript/scripts/make_docx.js" "$tmp\transcript_original.docx" "<VideoTitle> - Original" "<plain_text>"

# Chinese translation
node "~/.agents/skills/video-transcript/scripts/make_docx.js" "$tmp\transcript_zh.docx" "<VideoTitle> - 中文译稿" "<chinese_text>"

For Chinese video (one file):

node "~/.agents/skills/video-transcript/scripts/make_docx.js" "$tmp\transcript_zh.docx" "<VideoTitle> - 讲稿" "<plain_text>"

For long texts, write content to a temp .txt file first and pipe it:

$plain | Out-File -Encoding utf8 "$tmp\content.txt"
Get-Content "$tmp\content.txt" -Raw | node "~/.agents/skills/video-transcript/scripts/make_docx.js" "$tmp\transcript_original.docx" "<title>"

Step 6 — Deliver to user

Tell the user the output file paths clearly:

• "原文讲稿：C:\...\transcript_original.docx"
• "中文译稿：C:\...\transcript_zh.docx"

Optionally show a short preview (first 200 chars) of the extracted text so the user can verify quality.

Error handling

Notes

• Always clean up $tmp dir after delivering files, or leave it and tell the user the path
• The make_docx.js script resolves docx from global npm automatically
• For very long videos, translation may take a moment — let the user know