| name | yux-video-summary |
| description | Transform a video transcript file into a structured, organized summary with key points, timeline, and cleaned transcript. Use when the user has a transcript file (typically from yux-video-subtitle) and wants it summarized — e.g., "summarize this video", "organize the transcript", "video summary", "整理视频", "视频摘要", "内容总结". Supports bilingual output when video language differs from user language. Do NOT use for summarizing articles, documents, or non-video content. To convert the summary into a blog post, use yux-blog-writer. |
| allowed-tools | Read, Write, Glob, Grep |
Video Content Summary
Input: Transcript file path from $ARGUMENTS
Step 1: Load & Detect Languages
- Read transcript file (
.txt in current directory if not specified)
- Detect video_language: Chinese chars > 30% →
zh, else → en
- Detect user_language: From user's input message, same rule
- Output mode: Same language → SINGLE_FILE, Different → DUAL_FILE
- Inform user: "视频语言: [X], 用户语言: [Y], 模式: [Combined/Dual-file]"
Step 2: Identify Video Type
| Type | Indicators |
|---|
| Interview/Podcast | Multiple speakers, Q&A format |
| Talk/Lecture | Single speaker, educational |
| Tutorial/Review/News | Instructions, evaluation, reporting |
Step 3: Extract & Clean Content
3a. Extract Key Information
- Main topics and themes
- Key points and arguments
- Timeline highlights with timestamps
- Notable quotes
- Speaker identification (for interviews)
3b. Clean Transcript (Filler Word Removal)
English: um, uh, er, ah, you know, I mean, kind of, sort of, like, basically, actually, literally, right, okay
Chinese: 嗯, 啊, 呃, 哦, 那个, 这个, 就是, 然后, 所以说, 对吧, 是吧, 懂吗, 其实, 怎么说呢
Rules: Remove fillers, merge broken sentences, organize by topic/timeline sections
Step 4: Generate Output
Follow the templates in references/summary-templates.md for the exact output format:
- SINGLE_FILE mode: One combined file with summary + organized transcript
- DUAL_FILE mode: Separate summary file (user language) + transcript file (video language)
Step 5: Save & Report
SINGLE_FILE: Save <Title>-summary-<lang>.md, report path and mode
DUAL_FILE: Save both files, report:
Generated:
1. Summary: <Title>-summary-<user_lang>.md
2. Transcript: <Title>-transcript-<video_lang>.md
Mode: Dual-file
Step 6: Completion Report
=== Video Summary Generated ===
File(s): <Title>-summary-<lang>.md [+ transcript file if DUAL_FILE]
Mode: Single-file / Dual-file
Language: Video: <video_lang>, Summary: <user_lang>
Sections: Overview, Key Points (N topics), Timeline (N entries), Quotes (N)
Error Handling
| Situation | Action |
|---|
| Transcript file not found | Search for *.txt and *-transcript* in current directory. If none found, ask user for path |
| Transcript empty or under 100 words | Warn: "Transcript too short for meaningful summary." Ask user to confirm or provide a different file |
| Timestamp format inconsistent | Normalize to [HH:MM:SS] or [MM:SS] based on video length. If no timestamps, skip Timeline section |
| Language detection ambiguous | Default to English. Inform user of detected language and ask to confirm |
Quality Guidelines
- Concise and accurate
- Structured hierarchy
- Natural translations
- Preserve meaning when removing fillers