Video Insight

Video Insight

Analyzes YouTube videos or local media files to generate summaries, insights, and optionally Q&A highlights to reinforce key learning points.

Architecture

flowchart TB
    subgraph Main["Main Session"]
        SKILL[SKILL.md<br/>Orchestrator]
    end

    subgraph Agents["Subagents"]
        subgraph Haiku["Haiku Models"]
            QM[qa-generator<br/>Q&A Generation]
        end
        subgraph Sonnet["Sonnet Models"]
            TA[transcript-analyzer<br/>Transcript Analysis]
            DW[digest-writer<br/>Digest Writing]
            DR[deep-researcher<br/>Deep Research]
        end
    end

    SKILL --> TA
    SKILL --> DW
    SKILL --> QM
    SKILL --> DR

    TA -.->|Return Summary| SKILL
    DW -.->|Save Document| SKILL
    QM -.->|Q&A Section| SKILL
    DR -.->|Research Results| SKILL

    style Main fill:#f5f5f5,stroke:#333
    style Haiku fill:#e1f5fe,stroke:#0288d1
    style Sonnet fill:#fff3e0,stroke:#f57c00

Context Management: Main Session handles only orchestration. Long transcript processing is performed by Subagents to protect context.

Prerequisites

YouTube URL Processing:

• Requires yt-dlp (brew install yt-dlp)

Local File Processing:

• Requires whisper-cpp (brew install whisper-cpp)
• Requires ffmpeg (brew install ffmpeg)
• Whisper model download (automatic on first run)

Check dependencies: ./scripts/check_dependencies.sh

Supported Input Types

Workflow

Dependency Check (Before Starting)

CRITICAL: Check required dependencies before processing. If missing, show installation guide and stop immediately (do not retry).

For YouTube URL:

./scripts/check_dependencies.sh --youtube

For Local Media File:

./scripts/check_dependencies.sh --local

If exit code is 1 (missing dependencies):

1. Display the script output (shows missing tools and install commands)
2. Inform user: "Please install the required dependencies and try again."
3. Reference: references/prerequisites.md for detailed installation guide
4. Stop processing - do not attempt to continue or retry

Important: Do not repeatedly check or retry installation. The user must manually install dependencies and re-run the command.

Step 0: Detect Input Type

Determine if input is YouTube URL or local file:

YouTube URL Pattern:

^https?://(www\.)?(youtube\.com|youtu\.be)

Local File:

• Check file existence ([ -f "$INPUT" ])
• Determine type by extension

Branching:

• YouTube URL → Step 1A (YouTube metadata)
• Local media file → Step 1B (Local metadata)
• Subtitle file (srt/vtt) → Go directly to Step 3
• Invalid input → Error message

Step 1A: Extract YouTube Metadata

./scripts/extract_metadata.sh "{youtube_url}"

Extract from JSON result:

• title, channel, upload_date, duration, description
• chapters (if available)
• subtitles, automatic_captions (subtitle availability)

Step 1B: Extract Local File Metadata

./scripts/extract_local_metadata.sh "{file_path}"

Extract from JSON result:

• title (extracted from filename)
• duration (extracted with ffprobe)
• format (file format)
• source: "local" (local file indicator)

Step 2: Check Video Duration

If over 60 minutes, present options with AskUserQuestion:

question: "Video duration is {duration}. How would you like to proceed?"
options:
  - label: "Process entire video"
    description: "Process the full video (may take longer)"
  - label: "First 30 minutes only"
    description: "Process only the first 30 minutes"
  - label: "Cancel"
    description: "Cancel video processing"

Step 3: Extract Transcript

For YouTube URL:

./scripts/extract_transcript.sh "{youtube_url}" "/tmp/video-insight"

Subtitle priority: Korean manual > English manual > Korean auto > English auto

If no subtitles available, present options with AskUserQuestion:

question: "No subtitles found. How would you like to proceed?"
options:
  - label: "Summarize description only"
    description: "Create a brief summary from the video description"
  - label: "Cancel"
    description: "Cancel video processing"

For local media file:

./scripts/extract_local_transcript.sh "{file_path}" "/tmp/video-insight"

Convert speech-to-text with whisper.cpp (Korean default)

For existing subtitle file:

Copy srt/vtt file to /tmp/video-insight/ for use

Step 4: Analyze Transcript (Subagent)

Call transcript-analyzer (Sonnet):

Using Task tool:
- subagent_type: "transcript-analyzer"
- model: sonnet
- prompt: |
    Analyze the transcript file.

    - transcript_path: /tmp/video-insight/{title}.ko.srt
    - metadata: {metadata JSON}
    - language: ko

    Extract key content, timeline, and important quotes.

Result: Return only analysis results to main session (not entire transcript)

Step 5: Confirm Save Path

Confirm save path with AskUserQuestion:

question: "Where would you like to save the digest file?"
header: "Save path"
options:
  - label: "Default path"
    description: "outputs/video/{YYYY-MM-DD}__{title}.md"
  - label: "Current folder"
    description: "./{YYYY-MM-DD}__{title}.md"
  - label: "Custom path"
    description: "Specify a custom path"

If custom path selected: Request path input from user

Step 6: Write Digest (Subagent)

Call digest-writer (Sonnet):

Using Task tool:
- subagent_type: "digest-writer"
- model: sonnet
- prompt: |
    Write a digest document.

    - analysis_result: {Step 4 result}
    - metadata: {metadata}
    - output_path: {path confirmed in Step 5}
    - template_path: templates/video-insight.md

    Also perform proper noun correction and add background information.

Result: Markdown file saved confirmation message

Step 7: Additional Content Options

Present options with AskUserQuestion (multiSelect enabled):

question: "Would you like to add additional sections?"
header: "Options"
multiSelect: true
options:
  - label: "Q&A Section"
    description: "Add Q&A highlights (1-5 pairs based on content length)"
  - label: "Deep Research"
    description: "Conduct in-depth research with web search"
  - label: "Skip all"
    description: "Generate digest only without additional sections"

Step 8: Generate Additional Content (Parallel Execution)

Based on user selection, execute agents in parallel. Each agent returns content only (does not write to file).

If Q&A selected, call qa-generator (Haiku):

Using Task tool:
- subagent_type: "qa-generator"
- model: haiku
- prompt: |
    Generate Q&A section content.

    - digest_path: {file path from Step 6}
    - qa_patterns_path: references/qa-patterns.md

    Create 1-5 Q&A pairs (based on content length)
    highlighting key information from the video.
    Return the Q&A section content in markdown format
    (do not write to file).

If Deep Research selected, call deep-researcher (Sonnet):

Using Task tool:
- subagent_type: "deep-researcher"
- model: sonnet
- prompt: |
    Perform deep research.

    - digest_path: {file path from Step 6}
    - deep_research_reference: references/deep-research.md

    Collect related materials via web search.
    Return the Deep Research section content in markdown format
    (do not write to file).

Parallel Execution: If both options are selected, launch both Task tools in a single message for parallel execution.

Step 9: Append Results to Digest

After agents complete, append returned content to the digest file:

1. Read current digest content
2. Append Q&A section (if generated)
3. Append Deep Research section (if generated)
4. Write updated content to digest file

Step 10: Cleanup Temporary Files

After all tasks complete, confirm cleanup with AskUserQuestion:

question: "Would you like to clean up temporary subtitle files?"
header: "Cleanup"
options:
  - label: "Clean up"
    description: "Delete subtitle files in /tmp/video-insight/ folder"
  - label: "Keep"
    description: "Keep subtitle files for additional work"

If clean up selected:

rm -rf /tmp/video-insight/

Display message: "Temporary files have been cleaned up."

If keep selected:

Display file location:

Temporary subtitle files are kept at /tmp/video-insight/
Manual cleanup: rm -rf /tmp/video-insight/

Bundled Resources

Subagents

Error Handling

Context Management

Handled in Main Session:

• Metadata extraction (small JSON)
• User option selection
• Subagent orchestration
• Final result summary display

Handled in Subagent:

• Read/analyze long transcript (transcript-analyzer)
• Write detailed document (digest-writer)
• Generate Q&A section (qa-generator)
• Web search/deep research (deep-researcher)

Design Rationale

Multi-Agent Architecture: Reading long transcripts directly in main session quickly exhausts context. Processing in Subagents and returning only results protects main context.

Model Selection:

• Haiku: Simple/repetitive tasks (Q&A generation)
• Sonnet: Analysis/creative tasks (transcript analysis, digest writing, deep research)

Optional Q&A: Not all users want Q&A sections. Providing it as optional increases flexibility.

Separate Deep Research: Additional web search is an optional feature, incurring cost only when needed.