| name | mkt-insight-extractor |
| description | Extract key insights from video transcripts using the 5-type framework (Framework, Paradigm Shift, Warning, Diagnosis, Principle). USE WHEN user says 'extract insights', 'bóc insight', 'phân tích transcript', 'tìm insight từ video', 'analyze transcript insights', 'rút insight', 'insight từ transcript', 'extract key points from transcript'. |
Insight Extractor
Reads a video transcript and extracts the most valuable insights, categorized by the 5 insight types framework. Outputs a concise summary (<200 words) and a structured insights list — ready for Notion or any content system.
When to Use
- User provides a transcript file path and wants insights extracted
- Called by
trend-researcher agent after subtitle extraction, before Notion push
- User says "bóc insight", "extract insights", "phân tích transcript"
- User wants key takeaways from a video without reading the full transcript
Input
A transcript file path (plain text .txt file).
How to receive input:
- Direct file path:
research/youtube/trends/some-video/VIDEO_ID.txt
- Or raw transcript text pasted by user
Process
Step 1: Load Reference
Read references/insight-types.md to calibrate the 5 insight types and their detection signals.
Step 2: Read Transcript
Read the full transcript file. Understand the overall topic, structure, and teaching style.
Determine content type:
- Tutorial/How-to → prioritize Framework insights (step-by-step)
- Opinion/Commentary → prioritize Paradigm Shift and Principle insights
- Cautionary/Mistakes → prioritize Warning and Diagnosis insights
- Mixed → extract across all types
Step 3: Extract Insights
Scan the transcript for moments matching each insight type's signals (from reference).
Rules:
- Maximum 10 insights per transcript
- Each insight needs: type, title (one-liner), explanation (1-2 sentences)
- For tutorials: collapse step-by-step instructions into a single Framework insight with numbered sub-steps
- Skip filler, intros, CTAs, sponsor segments — focus on substance only
- If a moment could be two types, pick the stronger match
- Preserve the speaker's original phrasing when it's powerful
Step 4: Write Summary
Write a summary of the transcript in under 200 words.
Summary rules:
- What is this video about? (1 sentence)
- Who is it for? (1 sentence)
- What are the 2-3 biggest takeaways? (rest of summary)
- Write in Vietnamese if transcript is Vietnamese, English if English
- Be specific — no generic "this video talks about marketing"
Step 5: Format Output
Produce the final markdown output in this exact structure:
Output Format
## Summary
[Under 200 words summary]
## Key Insights
### 1. [Title] — `Framework`
[1-2 sentence explanation]
### 2. [Title] — `Paradigm Shift`
[1-2 sentence explanation]
### 3. [Title] — `Warning`
[1-2 sentence explanation]
...up to 10 insights
Type labels must be one of: Framework, Paradigm Shift, Warning, Diagnosis, Principle
Quality Criteria
What makes output GOOD vs BAD:
- GOOD: Each insight is specific, actionable, and clearly typed — a reader gets value without watching the video
- GOOD: Summary tells you exactly what the video covers and who benefits
- BAD: Generic insights like "the speaker talks about marketing strategies"
- BAD: Summary that just lists topics without specifics
- BAD: More than 10 insights (signal: you're not filtering enough)
- BAD: Including filler, self-promotion, or sponsor content as insights
References
references/insight-types.md — The 5 insight types with detection signals and examples. Load before Step 2.