Agent-Skills.md

Agent Skills: video-add-chapters

Add chapters to videos by transcribing, analyzing, and generating structured markdown documents with YouTube chapter markers. Optionally generate highlight videos.

UncategorizedID: jykim/ai4pkm-vault/video-add-chapters

Repository

jykim/ai4pkm-vault
jykim
33

Install this agent skill to your local

pnpm dlx add-skill https://github.com/jykim/ai4pkm-vault/tree/HEAD/_Settings_/Skills/video-add-chapters

Skill Files

Browse the full folder contents for video-add-chapters.

Download Skill

Loading file tree…

_Settings_/Skills/video-add-chapters/SKILL.md

Skill Metadata

Name
video-add-chapters
Description
Add chapters to videos by transcribing, analyzing, and generating structured markdown documents with YouTube chapter markers. Optionally generate highlight videos.

video-add-chapters

Transcribe videos using Whisper API, automatically detect chapter boundaries, and generate structured markdown documents with YouTube chapter markers. Optionally create highlight videos from selected segments.

When to Use This Skill

  • Transcribing long videos (20+ minutes) and splitting into chapters
  • Converting video transcripts into structured documentation
  • Generating YouTube chapter markers for video descriptions
  • Cleaning up raw transcripts into readable documents
  • Creating highlight videos from selected transcript segments

Example Results

Requirements

System

  • Python 3.7+
  • FFmpeg (for audio extraction)

Python Packages

pip install -r requirements.txt

Environment Variables

  • OPENAI_API_KEY - Required for Whisper API

How It Works

flowchart TB
    subgraph Auto[Automatic Processing]
        direction LR
        A[Video] --> B[Transcribe] --> C[Analyze] --> D[Generate] --> E[Clean]
    end
    subgraph Optional[Optional Review]
        F[Check & Adjust]
    end
    Auto -.-> Optional -.-> Auto

All steps run automatically without user intervention. Optional review step available if manual adjustment is needed.

Usage

Quick Start (Automated Pipeline)

# Run all steps automatically
python transcribe_video.py "video.mp4" --language ko --output-dir "./output"
python suggest_chapters.py "video.mp4" --output "chapters.json"
python generate_docs.py "video.mp4" --chapters "chapters.json" --output-dir "./output"
python clean_transcript.py "./output/merged_document.md" --backup

Step-by-Step Details

1. Transcribe Video

python transcribe_video.py "video.mp4" --language ko --output-dir "./output"

# Skip if transcript already exists (useful for workflow integration)
python transcribe_video.py "video.mp4" --skip-if-exists
  • Splits video into 15-minute chunks
  • Transcribes using Whisper API
  • Handles timestamp offsets automatically
  • Output: {video} - transcript.json

2. Detect Chapter Boundaries

python suggest_chapters.py "video.mp4" --output "chapters.json"
  • Analyzes transcript for topic transitions
  • Uses transition signal patterns (not pauses)
  • Output: chapters.json with suggested boundaries

3. Generate Documents

python generate_docs.py "video.mp4" --chapters "chapters.json" --output-dir "./output"
  • Creates individual chapter markdown files
  • Generates merged document and index
  • Outputs YouTube chapter markers

4. Clean Transcript

python clean_transcript.py "./output/merged_document.md" --backup
  • Removes filler words
  • Improves sentence structure
  • Enhances paragraph cohesion
  • Preserves timestamps and chapter boundaries

Optional: Manual Review

If chapter boundaries need adjustment:

  1. Edit chapters.json with corrected timestamps
  2. Re-run generate_docs.py to regenerate documents

Highlight Video Generation (Optional)

After completing the chapter workflow, you can create a highlight video by selecting specific segments from the transcript.

Method 1: Annotation in Transcript (Recommended)

Mark highlights directly in the transcript using <u> or == annotations.

flowchart LR
    A[Transcript] --> B["Add <u> or ==marks=="]
    B --> C[parse_highlight_annotations.py]
    C --> D[Highlight Script]
    D --> E[generate_highlights.py]
    E --> F[Highlight.mp4]

Quick Start:

# 1. Open transcript markdown and add annotations:
#    <u>text to highlight</u>  or  ==text to highlight==

# 2. Parse annotations to generate highlight script
python parse_highlight_annotations.py "transcript.md" --video "video.mp4"

# 3. (Optional) Edit the generated highlight script to add titles

# 4. Generate highlight video
python generate_highlights.py "transcript - highlight_script.md" --padding 0.5

Supported Annotation Formats:

  • <u>highlighted text</u> - HTML underline (visible in most editors)
  • ==highlighted text== - Markdown highlight (Obsidian compatible)

Features:

  • Consecutive highlighted segments are automatically merged
  • End times calculated from next segment start
  • Auto-generates titles from first few words

Method 2: Manual Script Editing

Export full transcript as editable script, then delete unwanted lines.

flowchart LR
    A[Transcript] --> B[Export Script]
    B --> C[User Deletes Lines]
    C --> D[Generate Video]
    D --> E[Highlight.mp4]

Quick Start:

# 1. Export editable highlight script
python export_highlight_script.py "video.mp4" \
  --transcript "./output/video - transcript.json"

# 2. Edit the script - delete unwanted lines (in any text editor)

# 3. Generate highlight video
python generate_highlights.py "./output/video - highlight_script.md"

Generate Highlight Video

python generate_highlights.py "highlight_script.md" --output "highlights.mp4" --padding 0.5 --title-duration 3
  • Parses [START-END] timestamps from script
  • Adds 0.5s padding before/after each segment to avoid mid-sentence cuts
  • Displays optional segment titles (yellow centered text, Korean font) for 3 seconds
  • Merges all segments into single video using FFmpeg
  • Output: {video} - highlights.mp4

Highlight Script Format

# Highlight Script: Video Title

**Source Video**: /path/to/video.mp4

---

[00:00:09-00:00:21] {Gemini CLI 설정} 우선은 Gemini에다가 제가 현재 커뮤니티 볼트가...

[01:46-02:15] {설치 완료} 네 설치가 된 것 같습니다. 볼트에서 확인해 볼까요.

[03:56-04:01] 아웃박스 커뮤니티 폴더에 질문을 생성을 했어요.

Format: [START-END] {Optional Title} Text content

  • Titles in {curly braces} are optional
  • If provided, title appears as yellow centered text overlay (144px, Korean font supported)

File Structure

video-add-chapters/
├── SKILL.md                    # This document
├── requirements.txt            # Python dependencies
├── transcribe_video.py         # Step 1: Video → Transcript
├── suggest_chapters.py         # Step 2: Chapter boundary detection
├── generate_docs.py            # Step 3: Document generation
├── clean_transcript.py         # Step 4: Transcript cleaning
├── parse_highlight_annotations.py  # Parse <u> and == annotations from transcript
├── export_highlight_script.py      # Export transcript as editable highlight script
├── generate_highlights.py          # Generate highlight video from script
├── templates/                  # Markdown templates
│   ├── chapter.md
│   ├── index.md
│   └── youtube_chapters.txt
└── examples/                   # Sample outputs
    ├── sample_chapter.md
    └── sample_youtube_chapters.txt

Troubleshooting

| Issue | Cause | Solution | |-------|-------|----------| | Chapter boundaries don't match content | Boundary set at keyword first mention | Use transition signal patterns for boundaries | | Merged document content mismatch | Manual updates missed in separate files | Update all related files when changing boundaries | | Transcript timing seems off | Misdiagnosed as offset issue | Verify: Whisper timestamps = video timestamps (no offset) | | Chapter content overlap | Boundary doesn't match content transition | Use end signals for endpoints, start signals for start points |

Verification Checklist

  1. [ ] Verify video timestamp at each chapter start
  2. [ ] Confirm next chapter content doesn't start before current chapter ends
  3. [ ] Check merged document body matches chapter boundaries
  4. [ ] Test YouTube link timestamps are accurate
  5. [ ] Verify original meaning is preserved after cleaning

Language Support

Currently optimized for Korean language with hardcoded transition patterns. Multi-language support: TBA.

Cost Estimate

Whisper API pricing: ~$0.006 per minute of audio

  • 1-hour video: ~$0.36
  • Processing is done in 15-minute chunks

Integration

Related Skills

  • video-full-process: Combined workflow with video-clean + chapter remapping
  • video-cleaning: Remove pauses and filler words
  • shorts-extraction: Extract chapters as short-form clips
  • youtube-transcript: Download YouTube video transcripts

Combined Workflow with video-clean

Use video-full-process skill for a unified pipeline that:

  1. Transcribes once (saving ~50% API costs)
  2. Removes pauses and filler words
  3. Remaps chapter timestamps to cleaned video
  4. Embeds chapters into final video
# From video-full-process skill directory
python process_video.py "video.mp4" --language ko

Transcript Reuse

The --skip-if-exists flag enables transcript reuse across skills:

# First transcription
python transcribe_video.py "video.mp4"

# Skip if already transcribed (reuses existing transcript)
python transcribe_video.py "video.mp4" --skip-if-exists

This prevents duplicate API calls when multiple skills need the same transcript

Reference: CHAPTERS Array Format

# Format: (start_seconds, title, description)
CHAPTERS = [
    (0, "Intro", "Introduction and welcome"),
    (98, "Setup", "Environment setup guide"),
    (420, "Main Content", "Core tutorial content"),
    # ... each chapter's start time, title, description
]