subtitle-extractor（字幕提取器）

Subtitle Extractor Skill

Extracts subtitles from video platforms in their native format. Supports Bilibili, YouTube, Xiaohongshu, Douyin, and local video files.

Scope of this skill: subtitle extraction only. Summarization, analysis, Q&A — all handled by the agent based on the user's actual request.

What It Does

1. Detect platform from URL
2. Extract native subtitles via yt-dlp (or Whisper transcription when no native subtitles exist)
3. Output: raw subtitle file path + video title/author
4. Agent saves subtitle to outputs/ and processes per user request

Dependencies

Agent must verify dependencies before calling the script. If any are missing, inform the user with the relevant install command.

yt-dlp — Required (always)

# Check
yt-dlp --version

# Install
pip install yt-dlp                    # all platforms (recommended)
brew install yt-dlp                   # macOS Homebrew
winget install yt-dlp.yt-dlp         # Windows WinGet
scoop install yt-dlp                  # Windows Scoop
conda install -c conda-forge yt-dlp  # Conda environments

# Upgrade existing install
pip install -U yt-dlp

ffmpeg — Required only for Whisper transcription

Only needed for Xiaohongshu, Douyin, local files, or Path B (Whisper transcription).

# Check
ffmpeg -version

# Install
brew install ffmpeg                   # macOS Homebrew
winget install Gyan.FFmpeg           # Windows WinGet
choco install ffmpeg                  # Windows Chocolatey
scoop install ffmpeg                  # Windows Scoop
apt install ffmpeg                    # Ubuntu / Debian
dnf install ffmpeg                    # Fedora / RHEL (may need RPM Fusion)
pacman -S ffmpeg                      # Arch Linux
snap install ffmpeg                   # Ubuntu Snap

Windows users: restart the terminal after installation for PATH to take effect. If winget is unavailable, download from ffmpeg.org and add the bin/ directory to system PATH.

faster-whisper — Required only for transcription platforms

Only needed for Xiaohongshu, Douyin, local files, or Path B (Whisper transcription).

# Check
python3 -c "from faster_whisper import WhisperModel; print('ok')"

# Install
pip install faster-whisper

# Configure model size (default: base)
export VIDEO_SUMMARY_WHISPER_MODEL=base   # tiny | base | small | medium | large

Note: Model files are downloaded automatically on first transcription run (~150MB for base). This may take a minute depending on network speed.

China network note: If auto-download fails (HuggingFace blocked), see Whisper model download failed in Troubleshooting.

Transcription time estimate (CPU, faster-whisper):

GPU accelerates transcription 5–15×. First run downloads the model (~150MB for base).

Cookie Configuration

Bilibili — Cookie Required

Bilibili requires a cookie file for all requests. The script auto-discovers cookie files in the skill directory only (same folder as subtitle-extractor.py and SKILL.md):

Any .txt file whose name contains bilibili will be picked up automatically — including the browser extension's default export format www.bilibili.com_netscape_<timestamp>.txt.

Place your cookie file in the skill directory. The agent does not need to locate or pass it manually (see Step 1b).

Xiaohongshu / Douyin — Manual

video-summary "https://www.xiaohongshu.com/explore/xxxxx" --cookies cookies.txt
# or
export VIDEO_SUMMARY_COOKIES=/path/to/cookies.txt

How to export cookies:

1. Install browser extension: "Cookie Editor（https://cookieeditor.org/）"
2. Log in to the platform
3. Export cookies to a .txt file（Netscape format）

Agent Workflow

EXECUTION ORDER — NON-NEGOTIABLE

Steps 1–4 in this skill MUST be completed in full before addressing any user request. The subtitle file MUST be saved to disk (Step 4) before the agent proceeds to summarization, translation, analysis, or any other task the user has asked for.

Treat Steps 1–4 as mandatory prerequisites, not optional helpers. Do not skip any step even if the user's final output format (e.g. a markdown file) appears to make it unnecessary.

Step 1 — Check Dependencies

yt-dlp --version

If the user requests Whisper transcription (keywords: "whisper转录" / "用whisper" / "transcribe" / "转录" / "语音转文字"), or the platform is Xiaohongshu, Douyin, or a local file, also check:

ffmpeg -version
python3 -c "from faster_whisper import WhisperModel; print('ok')"

If anything is missing, stop and tell the user which dependency to install (see Dependencies section).

Step 1b — Bilibili Cookie

The script auto-discovers any .txt file containing "bilibili" in the skill directory. Do not search for or pass the cookie file yourself.

Only act if the script exits with:

• 未找到 Bilibili Cookie 文件 → tell the user to place a cookie file in the skill directory
• Bilibili 412 错误：Cookie 已过期 → tell the user to re-export

To export: install "Cookie Editor (https://cookieeditor.org/)", log in to Bilibili, export Netscape format → place in skill directory → retry.

Step 2 — Extract Subtitles

Determine which path applies, then execute it completely before moving to Step 3.

Path A — Native subtitles

Use when: Bilibili or YouTube URL, and the user has not mentioned any transcription keyword.

Tell the user: "正在提取字幕..."

python subtitle-extractor.py "<url>"              # auto-detect language
python subtitle-extractor.py "<url>" --lang zh-CN  # force language

Parse the JSON from stdout. You now have all four fields needed for Step 3:

If the script exits non-zero: read stderr, report the error to the user, stop.

Path B — Whisper transcription

Use when: user mentions any transcription keyword, OR platform is Xiaohongshu or Douyin.

Transcription keyword takes priority over phrasing like "提取字幕" or "字幕原文" — those describe the desired output, not the method.

Call 1 — Download audio (skip for local files, go to Call 2 directly)

Tell the user: "正在下载音频，请稍候..."

python subtitle-extractor.py "<url>" --step download-audio

Parse the JSON from stdout and record these values:

If the script exits non-zero: read stderr, report the error to the user, stop.

Tell the user: "音频下载完成，开始 Whisper 转录（模型: base），请稍候..."

Call 2 — Transcribe

For URL input, use the audio_file recorded from Call 1:

python subtitle-extractor.py "<audio_file>" --step transcribe

For local file input (set title = filename, author = "local"):

python subtitle-extractor.py "<local_file_path>" --step transcribe

Parse the JSON from stdout and record:

Tell the user: "转录完成！"

If the script exits non-zero:

• Read stderr, report the error to the user, stop
• If stderr contains Whisper 模型下载失败: show the full error message verbatim — it contains the exact download directory and manual steps

Failure rule: Do not run yt-dlp, ffmpeg, or Whisper commands manually. Do not retry with different flags unless the error message explicitly says to.

Step 3 — Confirm Data

Verify you have collected all four values from the script outputs in Step 2:

Note: non-ASCII characters in JSON output appear as \uXXXX escapes — standard JSON parsing produces the correct decoded strings.

Step 4 — Save Subtitle to Outputs (REQUIRED — DO NOT SKIP)

Before answering the user, save the subtitle file to the session outputs directory.

Naming rule: {title前8字}_{author}.{原格式扩展名}

Steps:

1. Take title, keep the first 8 characters (Chinese and English each count as 1)
2. Replace unsafe filesystem characters / \ : * ? " < > | and spaces with _
3. Apply the same sanitization to author
4. Use the extension from subtitle_file path (.srt or .vtt)
5. Save to {outputs_dir}/{safe_title8}_{safe_author}.{ext}

Step 5 — Process and Respond

Read the subtitle file content and respond to the user's original request — summarize, analyze, translate, answer questions, etc. The subtitle content is in SRT or VTT format with timestamps; LLMs handle both directly.

Platform Notes

Supported URL Formats

YouTube: youtube.com/watch?v=... · youtu.be/...

Bilibili: bilibili.com/video/BV... · bilibili.com/video/av... · b23.tv/... (short link)

Xiaohongshu: xiaohongshu.com/explore/... · xhslink.com/... (short link)

Douyin: douyin.com/video/... · v.douyin.com/... (short link)

Script Reference

Usage:
    python subtitle-extractor.py <url|file> [options]

Steps (--step):
    download-audio    Download audio from URL → {"audio_file", "title", "author", "platform"}
    transcribe        Transcribe local audio/video file → {"subtitle_file"}
    (none)            Extract native subtitles (default path) → {"title", "author", "platform", "subtitle_file"}

Options:
    --step <name>     Pipeline step to run: download-audio | transcribe
    --lang <code>     Subtitle language code, default path only (default: auto)
    --cookies <file>  Cookie file for restricted content
    --help            Show help

Environment Variables:
    VIDEO_SUMMARY_COOKIES        Path to cookies file
    VIDEO_SUMMARY_WHISPER_MODEL  Whisper model size (default: base)

Troubleshooting

"yt-dlp: command not found"

pip install yt-dlp

"No subtitles found"

• The video may not have CC subtitles — use Path B (--step download-audio then --step transcribe) to force Whisper
• For Xiaohongshu/Douyin, transcription is always required (no native subtitles)
• Try --lang to specify a different language code

Bilibili 412 Precondition Failed

Cookie expired. Re-export:

1. Log in to Bilibili in browser
2. Use "Cookie Editor（https://cookieeditor.org/）" extension
3. Export（Netscape format）→ place in skill directory → retry

Bilibili: no zh-CN subtitle found

The script automatically falls back to ai-zh. If both fail, it lists all available subtitle codes. Use --lang <code> to specify one.

"Whisper not installed"

pip install faster-whisper

Whisper model download failed

The script tries hf-mirror.com then huggingface.co. If both fail (common in China), the script will print exact steps. Show the error message to the user verbatim — it contains the exact directory path and download URL.

Manual download (browser accessible in China):

1. Open: https://modelscope.cn/models/pkufool/faster-whisper-base/files
2. Download these 5 files: config.json model.bin tokenizer.json vocabulary.json preprocessor_config.json
3. Create the directory shown in the error message and place all 5 files there
4. Re-run the script — it auto-detects the local model, no download needed

For other model sizes (tiny/small/medium/large), change faster-whisper-base to faster-whisper-{size} in the ModelScope URL.

"ffmpeg not found" (during transcription)

See ffmpeg install commands in the Dependencies section above.

Video too long for Whisper

Use a smaller model:

export VIDEO_SUMMARY_WHISPER_MODEL=tiny

*Extract subtitles. Let the agent think.*