功能描述

Transcribe Bilibili videos to text with high accuracy using Whisper medium model. Use when the user provides a Bilibili video URL (BVxxxxx) and wants to: (1)...

使用说明 (SKILL.md)

Bilibili Transcript v2.2

Name: Bilibili Transcript
Author: 54lynnn

High-accuracy Bilibili video transcription with multi-language AI subtitle support.

Overview

This skill provides a complete transcription workflow for Bilibili videos:

Extract Video Metadata - Title, author, publish date, duration
Smart Subtitle Detection - Priority: CC subtitles → AI subtitles (multi-language) → Whisper transcription
Multi-language AI Subtitle Support - Auto-detects: ai-zh, ai-en, ai-ja, ai-es, ai-ar, ai-pt, ai-ko, ai-de, ai-fr
Browser Cookie Support - WSL Chromium or Windows Edge for member-only videos
Formatted Output - Saves as structured TXT file with metadata + summary placeholder + full transcript
Simplified Chinese - Automatically converts Traditional to Simplified Chinese

What's New in v2.2

✅ Fixed cookie detection - Now uses browser config directory instead of SQLite file (avoids encoding errors)
✅ One-stop solution - CC subtitles → AI subtitles → Whisper transcription, all in one script
✅ Better WSL support - Automatically detects WSL Chromium and Windows Edge cookies
✅ Smart fallback - Seamlessly switches between subtitle sources without user intervention

What's New in v2.1

✅ Improved cookie handling - Fixed UTF-8 encoding issues with snap Chromium
✅ Three-tier fallback - CC subtitles → AI subtitles → Whisper transcription
✅ Better error handling - Gracefully degrades when cookie sources fail

What's New in v2.0

✅ Multi-language AI subtitles - Supports 9 languages: Chinese, English, Japanese, Spanish, Arabic, Portuguese, Korean, German, French
✅ WSL Chromium support - Better cookie extraction than Windows Edge
✅ Correct subtitle download - Uses --write-subs --write-auto-subs combo
✅ Language auto-detection - Automatically finds available AI subtitle language

AI Subtitle Language Codes

Bilibili uses ai- prefix for AI-generated subtitles:

Code	Language	语言
`ai-zh`	Chinese	中文
`ai-en`	English	英文
`ai-ja`	Japanese	日文
`ai-es`	Spanish	西班牙文
`ai-ar`	Arabic	阿拉伯文
`ai-pt`	Portuguese	葡萄牙文
`ai-ko`	Korean	韩文
`ai-de`	German	德文
`ai-fr`	French	法文

Requirements

Hardware (Your Setup)

GPU: NVIDIA RTX 4070 Super (12GB VRAM) - ✅ Perfect for medium model
WSL Memory: 16GB (configured)
WSL CPU: 6 cores (configured)

Software

yt-dlp - Video/audio download
ffmpeg - Audio processing
whisper - Speech-to-text (local, no API key)
opencc - Traditional to Simplified Chinese conversion (optional)

Browser (for AI subtitles)

WSL Chromium (recommended) - Log in to Bilibili in WSL
Windows Edge - Alternative option

Workflow

Step 1: Run Transcription Script

./scripts/bilibili_transcript.sh "https://www.bilibili.com/video/BVxxxxx"

Priority order:

CC Subtitles (manual) - Fastest, highest accuracy
AI Subtitles (auto-generated) - Fast, good accuracy, multi-language
Whisper Transcription - Slowest, ~95% accuracy, works for all videos

Step 2: Generate Detailed Summary

After the script completes, read the generated TXT file and:

Read the full transcript (第二部分)
Generate a comprehensive summary (第一部分)
Save the updated file

Step 3: Present to User

In Discord, post:

Brief summary in message
Attach the TXT file for full content

Setup WSL Chromium Login

For best results with AI subtitles:

Start WSL Chromium:
```
chromium-browser &
```
Navigate to bilibili.com
Log in with your Bilibili account
Run the transcription script

The script will automatically use Chromium's cookies to access member-only AI subtitles.

Usage Examples

Example 1: Basic Transcription (Default Output)

./scripts/bilibili_transcript.sh "https://www.bilibili.com/video/BV1Z1wJzgEAj/"
# Output: workspace/Bilibili transcript/[VideoTitle]_BVxxxxx_transcript.txt

Example 2: Custom Output Directory

./scripts/bilibili_transcript.sh "https://www.bilibili.com/video/BV1Z1wJzgEAj/" ~/Documents

Notes

Model Selection

Your config: RTX 4070 Super 12GB + 16GB RAM + 6 cores
Default: medium model (~95% accuracy, balanced speed) ✅
Fallback: If GPU unavailable, automatically uses CPU (slower)

Accuracy Comparison

Source	Accuracy	Speed	Best For
CC Subtitles	100%	⚡ Instant	All videos with manual subtitles
AI Subtitles (ai-zh)	~90%	⚡ Instant	Chinese videos
AI Subtitles (ai-en)	~85%	⚡ Instant	English videos
Whisper medium	~95%	🐢 Slow	No subtitle videos

Default Output Directory

Location: workspace/Bilibili transcript/
Created automatically on first run
All transcript files organized in one place

File Naming

Output files are named: [VideoTitle]_[BVID]_transcript.txt

Special characters (including Chinese punctuation) are replaced with underscores
Title truncated to 50 characters
Example: 股票分红_是从左口袋掏右口袋吗_BV1ddzUYTE27_transcript.txt

Subtitle Priority

The script tries subtitles in this order:

Manual CC subtitles (zh-CN, zh-TW, en, ja, etc.)
AI subtitles (any available language: ai-zh, ai-en, ai-ja, etc.)
Whisper voice transcription (fallback)

This ensures fastest processing while maintaining high accuracy.

安全使用建议

What you should consider before installing or running this skill: - The script will attempt to read browser cookies (via yt-dlp --cookies-from-browser) to access member-only AI subtitles. That means local authentication cookies from Chromium/Edge may be read; if you are uncomfortable exposing those cookies to a third-party script, do not run it or run without cookie support. - The documentation and implementation disagree in a few places: README version (2.8.0) differs from registry (2.2.0); SKILL.md claims a default 'workspace/Bilibili transcript/' folder while the script defaults to /tmp; registry metadata lists no required binaries while the script requires yt-dlp, whisper, possibly ffmpeg and python3. Treat these as signs you should inspect the script before use. - The script probes the filesystem (e.g., lists /mnt/c/Users) to find Windows profile paths. If you run this on a multi-user or sensitive system, consider running it in a container or VM to avoid leaking local usernames or browser profiles. - If you trust the author and want full functionality, ensure yt-dlp, whisper and ffmpeg are installed from trusted sources. If you only want non-authenticated transcripts, run the script in no-cookie mode or remove the cookie-detection block. - If you are unsure, run the script in an isolated environment (container or throwaway VM), or manually step through the script to confirm behavior; the code is short and readable but performs sensitive local file access. Additional actions that would increase confidence: confirm the authoritative source/homepage and a stable matching version, or have the author update registry metadata to declare required binaries and the intended default output path.

功能分析

Type: OpenClaw Skill Name: bilibili-transcript Version: 2.2.0 The skill's primary script, `scripts/bilibili_transcript.sh`, automatically searches for and accesses sensitive browser cookie databases from both WSL Chromium and Windows Edge (via `/mnt/c/Users/`) to authenticate video downloads. While this behavior is documented in `SKILL.md` as a means to access member-only Bilibili content, the automated extraction of browser credentials represents a significant security risk. No evidence of intentional data exfiltration to external servers was found, but the capability to access host-system browser data warrants a suspicious classification.

能力评估

ℹ Purpose & Capability

Name/description match the actual script: it downloads subtitles or audio from Bilibili and transcribes with Whisper. However registry metadata says no required binaries while SKILL.md/README clearly require yt-dlp, whisper, ffmpeg and optionally opencc; README also lists a different version (2.8.0) than registry (2.2.0). SKILL.md claims default output folder 'workspace/Bilibili transcript/' but the script defaults OUTPUT_DIR to /tmp — documentation and implementation are inconsistent.

⚠ Instruction Scope

The runtime script probes local environment: it lists /mnt/c/Users to detect a Windows user and reads browser profile directories (WSL Chromium path and Windows Edge user data) to pass to yt-dlp's --cookies-from-browser. That means the script will attempt to read browser cookies (auth tokens) for member-only content. This is privacy-sensitive but is coherent with the stated goal of accessing member-only AI subtitles; still the script automatically searches for cookie locations without explicit interactive consent beyond 'log in in browser'.

✓ Install Mechanism

No install spec is present (instruction-only plus a shell script). No remote downloads or archive extraction are performed by the skill itself. The script relies on external binaries (yt-dlp, whisper, etc.) already on the host.

⚠ Credentials

No environment variables or external credentials are declared, which is good, but the script accesses local browser cookie stores via yt-dlp --cookies-from-browser and probes /mnt/c/Users. Those actions grant access to authentication cookies and reveal local usernames/profile paths. This access is proportionate to the stated need (member-only subtitles) but is sensitive and not explicitly reflected in registry 'required config paths' metadata.

✓ Persistence & Privilege

The skill is not always-enabled and does not request elevated privileges or modify other skills or system-wide configuration. It writes transcript files to an output directory (default /tmp or user-specified).

版本历史

v2.2.0

**bilibili-transcript v2.2.0 Changelog** - Fixed cookie detection: now uses the browser config directory to avoid encoding errors. - Streamlined workflow: automatically prioritizes CC subtitles, then AI subtitles, then Whisper transcription in a single script. - Improved WSL support: auto-detects both WSL Chromium and Windows Edge cookies. - Smarter fallback: seamlessly switches between available subtitle and transcription sources.

v2.16.0

Fix: Add fallback for empty duration display. Current file fixed to show '13分18秒'.

v2.15.0

Bug fixes: Fix empty duration display in TXT file. Improve filename handling to reduce UTF-8 truncation artifacts.

v2.14.0

Restructure TXT output into three clear sections: (1) Video Info with all metadata, (2) Summary generated from transcript, (3) Full transcript text.

v2.13.0

Improve duration format in filename: use 'XX时XX分XX秒' for videos >1 hour, 'XX分XX秒' for shorter videos.

v2.12.0

Fix: Use safer UTF-8 handling for filenames. Limit title to 40 chars and simplify duration format to avoid truncation artifacts.

v2.11.0

New filename format: Title_Author_Date_Duration_BVid.txt. Auto-generate video summary from transcript content. Discord preview shows brief summary.

v2.10.0

Fix: Improve filename sanitization with iconv UTF-8 handling. Extended max filename length to 80 chars.

v2.9.0

Lazy loading: Only check/install Whisper when voice transcription is actually needed. Most videos with subtitles work without Whisper installed.

v2.8.0

Complete Whisper model support: add tiny, small models. Now supports all 5 models (tiny/base/small/medium/large) with time estimation for each.

v2.7.0

Add interactive Whisper model selection with time estimation. Shows GPU info, estimated time for each model (base/medium/large), and lets user choose. Use -m or --model to pre-select, or -y for auto mode.

v2.6.0

Add language priority selection: use -l or --lang to specify AI subtitle language order. Example: -l en,zh for English first, then Chinese.

v2.5.0

Major update: Add multi-language AI subtitle support (ai-zh, ai-en, ai-ja, ai-es, ai-ar, ai-pt, ai-ko, ai-de, ai-fr). Add WSL Chromium cookie support.

v1.6.0

Add Windows Edge cookie support: automatically detects and uses Edge browser cookies for accessing member-only AI subtitles

v1.5.0

Add AI auto-subtitle support: prioritize Bilibili AI subtitles for faster transcription, fallback to manual subtitles, then Whisper

v1.4.0

Fix: Handle Chinese punctuation in filenames (？！，、；：''（）【】《》) to prevent encoding issues

v1.3.0

Fix: Use video title as filename, properly extract video metadata (title, author, publish date) from Bilibili

v1.2.0

Update: default Whisper model changed to medium, default output directory set to 'Bilibili transcript' folder

v2.1.0

v2.1.0: Switch to medium model for balanced speed/accuracy (~95%, 5GB VRAM, much faster than large)

v2.0.0

v2.0.0: Use Whisper large model with GPU, output formatted TXT file with metadata, auto convert to Simplified Chinese

元数据

Slug bilibili-transcript

版本 2.2.0

许可证 MIT-0

累计安装 5

当前安装数 5

历史版本数 22

常见问题

Bilibili Transcript 是什么？

Transcribe Bilibili videos to text with high accuracy using Whisper medium model. Use when the user provides a Bilibili video URL (BVxxxxx) and wants to: (1)... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 1057 次。

如何安装 Bilibili Transcript？

在 OpenClaw 或 Claude Code 对话框中运行命令「/install bilibili-transcript」即可一键安装，无需额外配置。

Bilibili Transcript 是免费的吗？

是的，Bilibili Transcript 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。

Bilibili Transcript 支持哪些平台？

Bilibili Transcript 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。

谁开发了 Bilibili Transcript？

由 54Lynnn（@54lynnn）开发并维护，当前版本 v2.2.0。

Bilibili Transcript