split-youtube-chapters

split-youtube-chapters Skill

Convert YouTube content into a chapter-based subpage structure.

Usage

/split-youtube-chapters <week/slug>

Example

/split-youtube-chapters week1/deep-dive-llms

Input/Output

Input

docs/week{N}/{slug}/eng/index.md - Original YouTube content including chapter structure
docs/week{N}/{slug}/kr/index.md - Full translation file (required when generating chapter files)

Output

src/content/readings.ts Modified:
- Convert the existing Reading into a isParent: true + children array
- Create each chapter as a ChildReading
docs/week{N}/{slug}/kr/{childSlug}.md - Per-chapter translation files (N)

Conversion example

Before conversion (single page)

'week1/deep-dive-llms': {
  slug: 'deep-dive-llms',
  week: 1,
  title: 'Deep Dive into LLMs like ChatGPT',
  titleKr: 'ChatGPT 같은 LLM 심층 분석',
  author: 'Andrej Karpathy',
  readTime: '약 3시간 31분',
  sourceUrl: 'https://www.youtube.com/watch?v=7xTGNNLPyMI',
  sourceTitle: 'YouTube - Deep Dive into LLMs like ChatGPT',
  published: true,
  sections: [
    { title: '1. Introduction (0:00)', content: '...' },
    { title: '2. Pretraining Data (1:00)', content: '...' },
    // ... 24개 섹션
  ],
  keyTakeaways: [...]
}

After conversion (parent + child structure)

'week1/deep-dive-llms': {
  slug: 'deep-dive-llms',
  week: 1,
  title: 'Deep Dive into LLMs like ChatGPT',
  titleKr: 'ChatGPT 같은 LLM 심층 분석',
  author: 'Andrej Karpathy',
  readTime: '약 3시간 31분',
  sourceUrl: 'https://www.youtube.com/watch?v=7xTGNNLPyMI',
  sourceTitle: 'YouTube - Deep Dive into LLMs like ChatGPT',
  published: true,
  isParent: true,
  // 부모 페이지의 sections는 개요 정보
  sections: [
    {
      title: '개요',
      content: 'Andrej Karpathy의 LLM 심층 분석 강의입니다...',
    }
  ],
  // YouTube 전용 필드 유지
  contentType: 'youtube',
  duration: '3:31:05',
  totalChapters: 24,
  tldr: '...',
  learningGoals: [...],
  chapterSummaries: [...],
  // 챕터별 자식 페이지
  children: [
    {
      slug: 'introduction',
      title: '1. Introduction',
      titleKr: '1. 소개',
      sourceUrl: 'https://www.youtube.com/watch?v=7xTGNNLPyMI&t=0s',
      published: true,      // 번역 파일 있으면 true
      hasMarkdown: true,    // 마크다운 동적 로딩 활성화
    },
    {
      slug: 'pretraining-data',
      title: '2. Pretraining Data',
      titleKr: '2. 사전학습 데이터',
      sourceUrl: 'https://www.youtube.com/watch?v=7xTGNNLPyMI&t=60s',
      published: true,
      hasMarkdown: true,
    },
    // ... 24개 챕터
  ]
}

Execution instructions

Step 1: Read the original markdown

1. docs/week{N}/{slug}/eng/index.md 파일 읽기
2. 파일이 없으면 에러: "원본 파일을 찾을 수 없습니다."
3. 프론트매터에서 메타데이터 추출:
   - source_url
   - chapters (또는 본문에서 파싱)

Step 2: Parse the chapter structure

Extract chapters from the Table of Contents or h2 sections of the original markdown:

## Table of Contents

1. [Introduction](#1-introduction) (0:00)
2. [Pretraining Data (Internet)](#2-pretraining-data-internet) (1:00)
3. [Tokenization](#3-tokenization) (7:47)
...

Information to extract from each chapter:

Number: 1, 2, 3, ...
English title: Introduction, Pretraining Data, Tokenization
Timestamp: 0:00, 1:00, 7:47

Step 3: Convert chapter title → slug

English title	slug
Introduction	`introduction`
Pretraining Data (Internet)	`pretraining-data`
Tokenization	`tokenization`
Neural Network I/O	`neural-network-io`
GPT-2: Training and Inference	`gpt2-training-inference`
Llama 3.1 Base Model Inference	`llama-31-base-model`
Hallucinations, Tool Use, Knowledge/Working Memory	`hallucinations-tool-use`
Reinforcement Learning from Human Feedback (RLHF)	`rlhf`

Slug generation rules:

Convert to lowercase English
Remove special characters (parentheses, colons, slashes, etc.)
Convert spaces to -Convert to
Remove consecutive - removal
If too long, use only key words (within about 30 characters)

Step 4: Translate chapter titles into Korean

English title	titleKr
Introduction	Introduction
Pretraining Data	Pretraining data
Tokenization	Tokenization
Neural Network I/O	Neural network input/output
Inference	Inference
Reinforcement Learning	Reinforcement learning
RLHF	RLHF (Reinforcement Learning from Human Feedback)

Korean translation rules:

Translate general technical terms into Korean
Keep proper nouns (GPT-2, Llama, DeepSeek-R1) as is
Keep abbreviations (RLHF, I/O) as is + explain in parentheses
Keep numbers as is: "1. Introduction", "2. Pretraining data"

Step 5: Convert timestamp → YouTube URL

타임스탬프 "1:20:32" → 초 단위 변환 → 4832초
YouTube URL: {sourceUrl}&t=4832s

Step 6: Update readings.ts

Find the existing Reading
Add/modify the following fields:
- isParent: true
- sections: Replace with overview information (remove existing sections)
- children: Add a per-chapter ChildReading array

// 기존 sections 대체
sections: [
  {
    title: '개요',
    content: `${author}의 ${titleKr} 강의입니다. 총 ${totalChapters}개의 챕터로 구성되어 있으며, 약 ${duration} 분량입니다.`,
  }
],

// children 추가
// 번역 파일이 있으면 published: true, hasMarkdown: true
// 번역 파일이 없으면 published: false, hasMarkdown: false
children: [
  {
    slug: 'introduction',
    title: '1. Introduction',
    titleKr: '1. 소개',
    sourceUrl: 'https://www.youtube.com/watch?v=...&t=0s',
    published: true,      // 번역 파일 있으면 true
    hasMarkdown: true,    // 번역 파일 있으면 true
  },
  // ...
]

Step 7: Check the readings.ts update result

After completing the readings.ts update, proceed to the next step.

Step 8: Generate per-chapter translation files (if not --skip-files)

Split each chapter from the full translation file and generate individual files.

8.1 Check the full translation file

파일 경로: docs/week{N}/{slug}/kr/index.md

파일이 없으면:
  ⚠ 전체 번역 파일이 없습니다.
  → 챕터별 파일 생성 건너뜀
  → readings.ts 업데이트만 완료

8.2 Parse the SUMMARY section

<!-- SUMMARY_START -->
## TL;DR
...

## 챕터별 요약

### 1. 소개 (0:00)
요약 텍스트...

**핵심 포인트:**
- 포인트1
- 포인트2

---

### 2. 사전학습 데이터 - 인터넷 (1:00)
...
<!-- SUMMARY_END -->

Parsing regex:

챕터 헤더: /^### (\d+)\.\s*(.+?)\s*\((\d+:\d+(?::\d+)?)\)/gm
  그룹: [1]=번호, [2]=한국어 제목, [3]=타임스탬프

핵심 포인트: /\*\*핵심 (?:포인트|개념):\*\*\n((?:- .+\n?)+)/g

Extracted data:

number: Chapter number
titleKr: Korean title (excluding number)
timestamp: Timestamp
summary: Summary text
keyPoints: Core points array

8.3 Parsing the FULL_TRANSLATION section

<!-- FULL_TRANSLATION_START -->
## 전체 번역

### 1. 소개

**요약**: ChatGPT와 같은...

[0:00] 안녕하세요, 여러분...
[0:30] 이 영상을 통해...

### 2. 사전학습 데이터 - 인터넷

**요약**: LLM 사전학습의...

[1:00] 도구를 준비하고...
<!-- FULL_TRANSLATION_END -->

Parsing regex:

챕터 구분: /^### (\d+)\.\s*(.+?)$/gm
타임스탬프 라인: /^\[(\d+:\d+(?::\d+)?)\]\s*(.+)$/gm

Extracted data:

number: Chapter number
fullTranslation: Full translation text including timestamps

8.4 Matching chapter data

Match the data extracted from SUMMARY and FULL_TRANSLATION by chapter number:

interface ChapterData {
  number: number;
  titleKr: string;       // SUMMARY에서
  timestamp: string;     // SUMMARY에서
  summary: string;       // SUMMARY에서
  keyPoints: string[];   // SUMMARY에서
  fullTranslation: string; // FULL_TRANSLATION에서
}

8.5 Creating per-chapter files

Create a file for each chapter:

Output path: docs/week{N}/{parentSlug}/kr/{childSlug}.md

Create directory: docs/week{N}/{parentSlug}/kr/ Create if the directory does not exist

File format:

---
title: "{N}. {English Title}"
titleKr: "{N}. {한국어 제목}"
chapter: {N}
timestamp: "{타임스탬프}"
sourceUrl: "{YouTube URL}&t={초}s"
translatedAt: "{YYYY-MM-DD}"
---

# {N}. {한국어 제목}

[영상 바로가기 ({타임스탬프})]({sourceUrl})

## 요약

{요약 텍스트}

**핵심 개념:**
{핵심 포인트 목록}

---

## 전체 번역

{타임스탬프 포함 번역 텍스트}

8.6 Handling existing files

파일이 이미 존재하는 경우:
  --overwrite 옵션: 덮어쓰기
  기본: 건너뜀 (⚠ 표시)

Step 9: Check chapter file creation results

Check the chapter file creation results:

On partial success:

✓ 챕터별 번역 파일 생성
  - 20개 파일 생성
  ⚠ 4개 파일 건너뜀 (이미 존재):
    - introduction.md
    - tokenization.md

No full translation file:

⚠ 전체 번역 파일이 없습니다: docs/week1/kr/deep-dive-llms.md
  챕터별 번역 파일 생성을 건너뜁니다.
  readings.ts 업데이트만 완료합니다.
  (children의 published: false, hasMarkdown: false)

다음 단계:
  /translate-reading week1/deep-dive-llms 실행 후 다시 실행하세요.

Detailed chapter parsing rules

Table of Contents format

## Table of Contents

1. [Introduction](#1-introduction) (0:00)
2. [Pretraining Data (Internet)](#2-pretraining-data-internet) (1:00)

Regex:

/^(\d+)\.\s*\[(.+?)\]\(#.+?\)\s*\((\d+:\d+(?::\d+)?)\)/gm

Group:

Number
Title
Timestamp

h2 section format

## 1. Introduction

**요약**: ChatGPT와 같은...

[0:00] hi everyone...