批量把抖音 / B 站 / 小宇宙 / YouTube / Vimeo / TikTok 等任意平台音视频转录为带语义切分的 Markdown 逐字稿（每视频一个独立 .md）。
特色：
- 5+0.5 阶段管线：fetch → transcribe（仅无字幕时）→ pack + 语言检测 → 3.5 translate（仅非中文时）→ boundaries → build
- 全平台支持：抖音 / B站 / 小宇宙 / YouTube + yt-dlp 1700+ 平台（Vimeo / TikTok / Twitter / X / Twitch / TED 等）
- 三层字幕探测（v4.1）：B站公共 API → yt-dlp → Whisper，平台有字幕时跳过 Whisper（节省 95% 能耗）
- 双语对比翻译：用户使用「双语转录」触发词时，非中文内容翻译为中文，输出「中文译文 + 英文原文 blockquote」段落对比格式
- 6 路由模型：lesson / explainer / conversation / demo / narrative / bulletin
- Light-plus 原则：保留推理 / 例子 / 数字 / 限定条件，非摘要
- 按语义边界切分，不机械每 60/90 秒切一段
- 错字自动修正（faster-whisper 在中文上的常见同音错字）
- 本地副作用：脚本会下载视频/音频流到
outputs/目录，调用 ffmpeg/yt-dlp/agent-browser 等外部进程 - 网络访问：B站公共 API、抖音 CDN、YouTube transcript API；YouTube 走
HTTPS_PROXY代理 - 资源占用：Whisper 转录会占用 GPU/CPU 资源（字幕可用时自动跳过）
- 清理方式：删除
outputs/目录可清理全部中间产物
pip install faster-whisper opencc-python-reimplemented youtube-transcript-api
pip install -U yt-dlp
winget install Gyan.FFmpeg
npm install -g agent-browser把链接写到 inputs.txt，每行一条（支持抖音分享文本、抖音 URL、B 站 URL 含 p=N 参数）：
7.94 复制打开抖音【教不会再杠Vivian的作品】第一讲-找到核心动词 https://v.douyin.com/xxx/
https://www.bilibili.com/video/BV1NK411D7pr?p=1
然后依次执行：
# 阶段 1：批量获取（字幕优先，无字幕时下载音频）
python scripts/01_fetch.py inputs.txt --output-dir outputs/batch
# 阶段 2：Whisper 转录（仅无字幕视频）
python scripts/02_transcribe.py --metadata outputs/batch/metadata.json --output-dir outputs/batch
# 阶段 3：打包 + 语言检测
python scripts/03_pack_transcript.py --transcript-dir outputs/batch/transcript --output-dir outputs/batch
# 阶段 3.5：翻译（仅非中文视频）
python scripts/03_5_translate.py prepare --output-dir outputs/batch/work
# 主对话循环：next-batch → AI 翻译 → apply（重复直到 done）
# 阶段 4：主对话读 work/boundary-review_N.md + takes_packed_N.md，写 scene-boundaries_N.json
# 阶段 5：输出 MD（自动识别双语）
python scripts/04_build_output.py --transcript-dir outputs/batch/transcript --boundaries-dir outputs/batch/work --output-dir outputs/batch最终输出在 outputs/batch/share/*.md。非中文视频在「双语转录」触发词下输出双语对比格式。
- SKILL.md — 主入口
- references/pipeline-details.md — 5+0.5 阶段管线 + 平台支持矩阵
- references/route-rules.md — 6 路由模型
- references/examples.md — 完整示例集
- references/troubleshooting.md — 故障排查 + 首次使用配置
MIT-0
Batch transcribe Douyin / Bilibili / Xiaoyuzhou / YouTube / Vimeo / TikTok and other platform audio/video into semantically segmented Markdown transcripts (one .md per video).
Highlights:
- 5+0.5 stage pipeline: fetch → transcribe (only when no subtitles) → pack + language detection → 3.5 translate (only for non-Chinese) → boundaries → build
- Full platform support: Douyin / Bilibili / Xiaoyuzhou / YouTube + yt-dlp 1700+ platforms (Vimeo / TikTok / Twitter / X / Twitch / TED, etc.)
- Three-layer subtitle detection (v4.1): Bilibili public API → yt-dlp → Whisper; skip Whisper when platform subtitles available (saves 95% energy)
- Bilingual contrast translation: when user triggers with "双语转录" (bilingual transcribe), non-Chinese content is translated to Chinese, output as "Chinese translation + English original blockquote" paragraph contrast format
- 6-route model: lesson / explainer / conversation / demo / narrative / bulletin
- Light-plus principle: keep reasoning / examples / numbers / qualifiers, not a summary
- Semantic boundary segmentation, not mechanical 60/90s cuts
- Auto error correction (common faster-whisper homophone errors in Chinese)
- Local side effects: Scripts download video/audio streams to
outputs/directory, invoke ffmpeg/yt-dlp/agent-browser external processes - Network access: Bilibili public API, Douyin CDN, YouTube transcript API; YouTube uses
HTTPS_PROXY - Resource usage: Whisper transcription consumes GPU/CPU resources (auto-skipped when subtitles available)
- Cleanup: Delete the
outputs/directory to clean all intermediate artifacts
pip install faster-whisper opencc-python-reimplemented youtube-transcript-api
pip install -U yt-dlp
winget install Gyan.FFmpeg
npm install -g agent-browserWrite links to inputs.txt, one per line (supports Douyin share text, Douyin URLs, Bilibili URLs with p=N parameter):
https://www.bilibili.com/video/BV1NK411D7pr?p=1
https://www.youtube.com/watch?v=xxxxx
Then run:
# Stage 1: Batch fetch (subtitle-first, fallback to audio download)
python scripts/01_fetch.py inputs.txt --output-dir outputs/batch
# Stage 2: Whisper transcription (only for videos without subtitles)
python scripts/02_transcribe.py --metadata outputs/batch/metadata.json --output-dir outputs/batch
# Stage 3: Pack + language detection
python scripts/03_pack_transcript.py --transcript-dir outputs/batch/transcript --output-dir outputs/batch
# Stage 3.5: Translation (only for non-Chinese videos)
python scripts/03_5_translate.py prepare --output-dir outputs/batch/work
# Main conversation loop: next-batch → AI translate → apply (repeat until done)
# Stage 4: Main conversation reads work/boundary-review_N.md + takes_packed_N.md, writes scene-boundaries_N.json
# Stage 5: Output MD (auto-detect bilingual)
python scripts/04_build_output.py --transcript-dir outputs/batch/transcript --boundaries-dir outputs/batch/work --output-dir outputs/batchFinal output in outputs/batch/share/*.md. Non-Chinese videos output bilingual contrast format under the "双语转录" (bilingual transcribe) trigger.
- SKILL.md — Main entry
- references/pipeline-details.md — 5+0.5 stage pipeline + platform matrix
- references/route-rules.md — 6-route model
- references/examples.md — Full example set
- references/troubleshooting.md — Troubleshooting + first-use config
MIT-0