In a nutshell
Throw in a video, get out a complete set of content products: transcript + insight summary + short video clips + publishable article.
Pipeline overview
输入: 视频文件 (直播回放/播客/会议录像)
│
▼
┌─────────────────────────────────────────────┐
│ Stage 1: video-to-text │
│ 转写 → 带时间戳的文字稿 + JSON │
│ (去口吃的基础数据也在这里产生) │
└──────────────────┬──────────────────────────┘
│
┌──────────┴──────────┐
▼ ▼
┌───────────────┐ ┌─────────────────────┐
│ Stage 2: │ │ Stage 3: │
│ insight- │ │ video-clipper │
│ extractor │ │ 按观点切片 → │
│ 提炼观点/金句 │ │ 去静音/去口吃 → │
│ /争议点 │ │ 短视频成品 │
└───────┬───────┘ └─────────────────────┘
│
▼
┌───────────────────────────────────────────┐
│ Stage 4: article-forge │
│ 观点摘要 + 原始文稿 → 可发布文章 │
│ (博客/知乎/公众号) │
└───────────────────────────────────────────┘
输出目录: workspace/pipeline/<project-name>/
├── transcript.txt # 完整文字稿
├── transcript.json # 带时间戳 JSON
├── insights.md # 观点摘要
├── clips/ # 短视频切片
│ ├── 01-xxx.mp4
│ ├── 02-xxx.mp4
│ └── ...
└── articles/ # 生成的文章
├── blog-xxx.md
└── zhihu-xxx.md
Execution flow
Input parameters
The user provides:
- Video file path(required)
- Project name(optional, defaults to generated from the filename)
- Target outputs(optional, defaults to all):
transcript— transcript onlyinsights— transcript + insightsclips— transcript + clipsarticles— transcript + insights + articleall— all (default)
- Article platform(optional): blog / zhihu / wechat
- Number of clips(optional, defaults to 5-8 clips)
Stage 1: Transcription (video-to-text)
- Read
video-to-textSKILL.md - Create the project directory:
workspace/pipeline/<project-name>/ - Run the transcription script in the background with nohup:
nohup python3 {video-to-text-skillDir}/scripts/transcribe.py \ /path/to/video.mp4 \ --output-dir workspace/pipeline/<project-name>/ \ --output-name transcript \ --diarize \ > /tmp/pipeline-transcribe.log 2>&1 & - Wait for completion (use process poll or check the output files)
- Output:
transcript.txt+transcript.json
⚠️ This is the most time-consuming stage, a 30-minute video takes about 10-20 minutes to transcribe. Subsequent stages are all fast.
Stage 2: Insight extraction (insight-extractor)
- Read
insight-extractorSKILL.md - Input
transcript.txt - Extract following the insight-extractor process:
- Topic boundary identification
- Core argument extraction
- Key quote annotation (preserve timestamps)
- Controversial point annotation
- Output:
insights.md
Key: The timestamps of key quotes and insights must be precise; Stage 3 clipping needs them.
Stage 3: Video clipping (video-clipper)
Can run in parallel with Stage 2 (both depend only on Stage 1's output).
- Read
video-clipperSKILL.md - Based on
insights.mdinsights +transcript.jsontimestamps to locate clip boundaries - Execute the four-stage clipping process:
- Use
batch-clip-v4.shfor batch clipping (remove silence + remove stutters + crossfade) - Use
batch-postcheck.shfor secondary quality check - If necessary, use
iterate-until-clean.pyfor iterative fixes
- Use
- Output:
clips/01-xxx.mp4,clips/02-xxx.mp4, ...
Optimization: Prioritize clipping the segments marked as "key quotes" and "controversial points" in insights—these have the most viral potential as short videos.
Stage 4: Article generation (article-forge)
- Read
article-forgeSKILL.md - Input
insights.md+transcript.txt - Select the genre and style according to the target platform
- Generate the article, strictly removing the AI flavor
- Output:
articles/blog-xxx.mdorarticles/zhihu-xxx.md
Optional: If the user requests publishing, call ZhiForge's publishing process.
Parallel strategy
时间线:
─────────────────────────────────────────────────>
Stage 1 (转写) Stage 2 (观点) Stage 4 (文章)
████████████████ ████████ ████████
Stage 3 (切片)
████████████████
- Stage 2 and Stage 3 canrun in parallel(run separately with sessions_spawn)
- Stage 4 depends on Stage 2's output, serial
- Stage 1 is the slowest, accounting for 60-70% of total time
Sub-agent orchestration
Recommended to use sessions_spawn parallelization:
主 agent:
1. 启动 Stage 1(转写,等待完成)
2. spawn 子 agent A → Stage 2(观点提炼)
3. spawn 子 agent B → Stage 3(视频切片)
4. 等 A 完成 → 启动 Stage 4(文章生成)
5. 等 B 完成 → 汇总输出
Output summary template
After everything is complete, report to the user:
🎬 视频内容处理完成
📹 源文件:xxx.mp4 (时长 XX:XX)
📁 项目目录:workspace/pipeline/<name>/
📝 文字稿:transcript.txt (XXXX 字)
💡 观点提炼:insights.md
- X 个核心观点
- X 条金句
- X 个争议点
🎞️ 短视频切片:X 条
- 01-xxx.mp4 (XX:XX) — 主题
- 02-xxx.mp4 (XX:XX) — 主题
- ...
📄 文章:X 篇
- blog-xxx.md (XXXX 字) — 标题
- zhihu-xxx.md (XXXX 字) — 标题
Quick trigger
If the user says any of the following, trigger this skill:
- "Process this video"
- "Full video pipeline"
- "Livestream replay processing"
- "Turn this video into content"
- "video pipeline"
Execute by default all(all outputs), unless the user specifies only a certain stage.
Notes
- Stage 1 transcriptionMust use nohup in the background, otherwise it will time out
- If the filename contains Chinese, create an English symlink
- Slicing depends on word_segments in transcript.json, not transcript.txt
- Article generation depends on insights.md; you cannot skip Stage 2 and go directly to Stage 4
- Store all outputs in
workspace/pipeline/<project-name>/, not /tmp/ - After using the browser, you must
browser stop