Youtube Learn

YouTube-Learn - Belief Archaeology

> Philosophy: You are an archaeologist, not a summarizer. Transcript = surface layer. > What we need to learn are the beliefs and worldviews that produce those phenomena.

Dependencies

• youtube-transcript-api (pip install youtube-transcript-api) - v1.2.4+
• Script: scripts/fetch_transcript.py

Configuration

Customize these paths in your environment before using:

STORAGE_ROOT = ~/youtube-learn/              # Where video analysis folders are stored
KNOWLEDGE_INDEX = ~/knowledge/index.md       # Your knowledge base index (for Vault Check, optional)

When to Use

User shares a YouTube link and wants deep analysis - not just a summary. Trigger: "youtube-learn [URL]", "deep learn this video", "analyze worldview", "belief archaeology"

Procedure

Phase 1 - Data Collection

1. Create video folder: /YYYY-MM-DD_short-video-name/
2. Use run_command to execute the transcript script:

`` python "/scripts/fetch_transcript.py" "/transcript.txt" ``

1. Script auto-handles: extract video ID from URL, fetch title + channel, get transcript
2. Language priority: Vietnamese -> English -> any available language
3. Read saved transcript.txt file using view_file for analysis
4. If script returns error (no transcript) -> inform user, ask if they want to analyze from description
5. DO NOT use browser_subagent - API script is faster and more stable
6. Transcript MUST be saved in the video folder (not in scratch or artifact dirs)

Phase 2 - Speaker Portrait

1. Use search_web to research speaker: identity, achievements, signature quotes
2. Prioritize LESSER-KNOWN facts - don't repeat Wikipedia's first paragraph
3. Output: name, top 5 achievements, what makes them different, quote + translation

Phase 3 - Belief Archaeology ⚡ (CORE - most important)

Select maximum 3 worldviews that matter most. Analyze each with 5 dimensions:

| Dimension | Content | |-----------|---------| | 📍 Phenomenon | Original quote + specific context | | 💡 Hidden Belief | 1 memorable sentence LIKE A PROVERB - what must be true in their mind for them to say this | | ✅ Works When | 2 concrete examples anyone can understand | | ❌ Fails When | 1-2 cases where this worldview breaks down | | 🎯 Apply To Me | 1 most specific actionable suggestion |

> RULE: "Hidden Belief" MUST read like a proverb, NOT complex philosophy.

Phase 4 - Synthesis

• Table of 3 core beliefs
• Compare vs mainstream: how do they see the world differently from most people?
• Internal tension: do their worldviews contradict each other?

Phase 5 - Vault Check (Cross-reference Knowledge Base)

1. Read your knowledge index file (configured in KNOWLEDGE_INDEX)
2. Use grep_search to find existing knowledge items related to the extracted worldviews
3. Classify: 🟢 Reinforces / 🔴 Contradicts / 🟡 Extends / 🔵 Completely new
4. Skip this phase if no knowledge index is configured

Phase 6 - Reflection Questions

• Provide 3 questions for the user to internalize
• Ask: which worldview resonates most? What do you want to save to your Worldview Library?

Phase 7 - Atomize (Storage)

Storage root: / Naming convention: YYYY-MM-DD_short-video-name/

Each video creates 1 dedicated folder containing ALL related data:

youtube-learn/
├── _INDEX.md                              /[video-folder]/speaker.md` - no user approval needed. Use template: `references/speaker_profile_template.md`
- **7B - Worldview Library:** MUST ASK user first -> only save the worldviews user confirms to `/[video-folder]/worldviews/worldview-[name].md`. Use template: `references/worldview_template.md`
- **7C - Update _INDEX.md:** AUTO-ADD new entry to all 3 tables (Videos, Speakers, Worldviews) in `/_INDEX.md`

### Phase 8 - QA Checklist
Check 9 criteria. Fail ANY item -> go back to that phase, DO NOT skip:
1. Transcript is complete (not cut off mid-way)?
2. Speaker Portrait has verified sources (no fabrication)?
3. Each worldview has all 5 analysis dimensions?
4. "Hidden Belief" reads like a proverb (not academic)?
5. "Works When" has concrete examples (not abstract)?
6. "Apply To Me" is actionable (can be done immediately)?
7. Vault Check cross-referenced with existing knowledge?
8. Speaker Profile has been saved/updated?
9. `_INDEX.md` has been updated (3 tables: Videos, Speakers, Worldviews)?

## Pitfalls
- **DO NOT summarize the video** - the goal is excavating worldviews, not creating a summary
- **DO NOT write verbose "Hidden Beliefs"** - must be concise like folk proverbs
- **DO NOT skip QA** - each failed criterion requires going back to the corresponding phase
- **DO NOT save Worldviews without user confirmation** (Phase 7B)
- **Speaker Profile CAN be auto-saved** (Phase 7A) - it's factual data
- **DO NOT scatter files** - all output for 1 video MUST be in the same folder
- **DO NOT forget to update `_INDEX.md`** - it's the quick-lookup map when you have many videos

### ⛔ CRITICAL RULE: SAVE TO PERSISTENT STORAGE
- **ALWAYS** save analysis to your configured `STORAGE_ROOT`, not temporary/ephemeral directories
- Reason: Files in temp directories are lost when starting a new chat session

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [dotanminh](https://github.com/dotanminh)
- **Source:** [dotanminh/youtube-learn-skill](https://github.com/dotanminh/youtube-learn-skill)
- **License:** MIT

Install and usage instructions live in the source repository linked above.