iopho-analyzing-videos — Video → .storyboard.md Reverse Engineering

Analyze a video file and generate a detailed .storyboard.md file — a human+AI readable "source code" representation of the video, including scene-by-scene breakdowns, visual descriptions, audio transcripts, and extracted keyframes.

Prerequisites

!python3 -c "import google.generativeai; print('google-generativeai: ✓')" 2>/dev/null || echo "google-generativeai: ✗ — install: pip install google-generativeai"

!which ffmpeg 2>/dev/null && echo "ffmpeg: ✓ (needed for frame extraction)" || echo "ffmpeg: ✗ — install: brew install ffmpeg"

!echo "GEMINI_API_KEY: $([ -n \"$GEMINI_API_KEY\" ] && echo '✓ set' || echo '✗ — set via: export GEMINI_API_KEY=your-key')"

Required: google-generativeai + ffmpeg + GEMINI_API_KEY environment variable.

Get a Gemini API key at https://aistudio.google.com/apikey (free tier available).

What It Produces

Given a video file, this skill generates:

output.storyboard.md    ← YAML frontmatter + scene-by-scene markdown
frames/
  scene-001.jpg         ← Keyframe at midpoint of scene 1
  scene-002.jpg         ← Keyframe at midpoint of scene 2
  ...

.storyboard.md Format

---
title: "Video Title"
duration_seconds: 19
resolution: "1920x1080"
style:
  visual_style: "clean, modern, minimalist"
  color_palette: ["#E6E6FA", "#9400D3", "#00BFFF"]
audio:
  has_voiceover: true
  voiceover_language: "en"
content:
  type: "product_demo"
  framework: "PAS"
  key_message: "..."
---

### Scene 1: Hook (0:00 – 0:02, 2s)
**Thumbnail**: ![scene-001](./frames/scene-001.jpg)

#### Visual
- **Shot Type**: wide
- **Camera Movement**: static
- **Subject**: ...
- **Text On Screen**: "..."

#### Audio
- **Voiceover**: "exact transcript"
- **Music**: upbeat, electronic

#### Analysis
- **Narrative Purpose**: ...
- **Sales Element**: problem
- **Transition Out**: cut to →

Each scene includes: Visual (shot type, camera, subject, text, graphics), Audio (voiceover, music, SFX), and Analysis (narrative purpose, viewer psychology, sales element).

Quick Reference

Analyze a local video:

python3 scripts/video_to_storyboard.py "$VIDEO_PATH" "$OUTPUT_PATH"

Analyze without frame extraction:

python3 scripts/video_to_storyboard.py "$VIDEO_PATH" "$OUTPUT_PATH" --no-frames

Use a different Gemini model:

GEMINI_MODEL=gemini-2.5-flash python3 scripts/video_to_storyboard.py "$VIDEO_PATH"

Execute

Run the analysis pipeline:

# From $ARGUMENTS: <video_path> [output_path] [--no-frames] [--model MODEL]
#
# Default output: same directory as video, with .storyboard.md extension
# Default model: gemini-2.0-flash
#
# The agent should:
# 1. Verify video file exists
# 2. Check GEMINI_API_KEY is set
# 3. Run the analysis script
# 4. Report: output path, scene count, frame count, token usage

python3 scripts/video_to_storyboard.py "$VIDEO_PATH" "$OUTPUT_PATH"

Pipeline Details

The analysis pipeline works in 4 stages:

Stage 1: Upload
  Video file → Gemini File API (supports up to ~1GB)
  Wait for server-side processing

Stage 2: Analyze
  Gemini Flash processes video natively (no frame extraction needed)
  Generates complete .storyboard.md with scene timestamps
  ~$0.001 per short video (< 30s)

Stage 3: Extract Frames
  Parse scene timestamps from generated markdown
  ffmpeg screenshots at each scene midpoint
  Insert thumbnail references into markdown

Stage 4: Output
  Write .storyboard.md file
  Write frames/ directory with keyframe JPGs
  Clean up uploaded file from Gemini

Script Setup

If the analysis script doesn't exist yet, create it:

# Check if script exists
ls scripts/video_to_storyboard.py 2>/dev/null || echo "Script not found — see references/video_to_storyboard.py for the full implementation"

The script requires these Python packages:

pip install google-generativeai

Cost Estimates

Video Length	Gemini Flash Cost	Notes
< 30s	~$0.001	Short ads, demos
1-5 min	~$0.01	Explainers, tutorials
5-15 min	~$0.03	Long-form content
15-60 min	~$0.06-0.15	Full presentations

Costs based on Gemini 2.0 Flash pricing. Gemini 2.5 Flash may differ.

Combine with Other iopho Skills

Full pipeline — search, download, analyze:

/iopho-searching-videos SaaS explainer under 30s --limit 5
/iopho-getting-videos https://youtube.com/watch?v=BEST_MATCH --mode all --output temp/
/iopho-analyzing-videos temp/video.mp4 temp/output.storyboard.md

Analyze + regenerate with Remotion: After generating .storyboard.md, use it as a blueprint to recreate the video with Remotion or other video generation tools.

Tips

• Gemini Flash processes video natively — no need to pre-extract frames
• The .info.json from iopho-getting-videos is auto-detected for source URL and title
• Short videos (< 30s) produce the most accurate scene detection
• For longer videos, consider splitting first with ffmpeg -ss START -to END
• The generated .storyboard.md is both human-readable and machine-parseable
• Use --no-frames for faster processing when you don't need keyframe images