video-analyst 1.5.0

Analyze videos and produce AI reproduction plans using Veo 3 and Nano Banana 2

video-analyst

Analyze videos from YouTube or TikTok and produce structured AI reproduction plans with Veo 3 video prompts, Nano Banana 2 image prompts, and natural voiceover text.

Downloads a video → analyzes it with Gemini 2.5 Flash → outputs a scene-by-scene reproduction plan with generation prompts, character reference sheets, and humanized voiceover in any language.

Install

As a CLI tool

# Clone and install
git clone https://github.com/getvrex/video-analyst.git
cd video-analyst
uv venv && uv pip install -e .

# Or with pip
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

As a Claude Code skill

# One-liner install
npx @getvrex/video-analyst install

# Or manually
git clone https://github.com/getvrex/video-analyst.git ~/.claude/skills/video-analyst
cd ~/.claude/skills/video-analyst && uv venv && uv pip install -e .

Via npm (for Claude Code skill)

npm install -g @getvrex/video-analyst

Setup

export GEMINI_API_KEY=your_key_here

Usage

# Full analysis in English (JSON output)
video-analyst analyze "https://youtube.com/watch?v=abc"

# Summary in Vietnamese, markdown output
video-analyst analyze "https://www.tiktok.com/@user/video/123" -m summary -l vi -f markdown

# With a visual style
video-analyst analyze "https://youtu.be/xyz" --style cyberpunk -l ja -f markdown

# Save to file
video-analyst analyze "https://youtube.com/shorts/abc" -o plan.json

# List available styles
video-analyst styles

Options

Flag	Short	Default	Description
--mode	-m	full	summary (condensed) or full (comprehensive)
--lang	-l	en	Target language for voiceover (en, vi, ja, ko, zh, es, ...)
--style	-s	realistic	Visual style preset for all prompts
--format	-f	json	Output format: json or markdown
--output	-o	stdout	Save output to file
--keep-video			Keep downloaded video after analysis
--model		gemini-2.5-flash	Override Gemini model
--verbose	-v		Show detailed progress

Claude Code

If installed as a skill, use it directly in Claude Code:

/video-analyst https://youtube.com/watch?v=abc
/video-analyst https://www.tiktok.com/@user/video/123 --summary --lang vi

Visual Styles

18 predefined styles that apply consistently across all generation prompts:

Style	Description
realistic	Ultra photorealistic — 35mm film, natural grain, no AI look (default)
cinematic	Hollywood blockbuster — anamorphic, dramatic lighting, teal/orange
ghibli	Studio Ghibli — hand-painted, lush backgrounds, whimsical
wes-anderson	Symmetrical, pastel palette, meticulous production design
noir	Film noir — high contrast B&W, dramatic shadows, 1940s
cyberpunk	Neon-lit dystopia — magenta/cyan, rain-slicked streets
vintage-film	1970s analog — heavy grain, faded colors, Kodak film stock
anime	Modern anime — sharp linework, vivid colors, dynamic angles
watercolor	Soft painted — visible brush strokes, flowing colors, dreamy
documentary	Raw handheld — 16mm film, natural light, authentic
kurzgesagt	Flat design motion graphics — vibrant vector, rounded shapes, bold colors
synthwave	80s retro-futuristic — neon grids, chrome, sunset gradients, laser lines
comic-book	Pop art — bold ink outlines, halftone dots, vivid primary colors
pixel-art	Retro 16-bit gaming — chunky pixels, limited palette, nostalgic
paper-cutout	Paper craft — layered cardstock, hand-cut edges, handmade charm
low-poly	Faceted 3D geometry — crystalline surfaces, vibrant gradients, minimal
claymation	Stop-motion clay — sculpted figures, fingerprint textures, warm lighting
vaporwave	90s surreal digital — pastel gradients, glitch effects, marble busts

Output Structure

VideoReproductionPlan
├── title                    # in target language
├── description              # video concept and hook
├── metadata_tags[]          # bilingual hashtags
├── viral_structure_notes    # hook analysis, content arc
├── characters[]
│   ├── character_name
│   ├── character_description    # identical across all scenes
│   └── t2i_reference_prompt     # Nano Banana 2, plain background, multi-angle
├── scenes[]
│   ├── scene_number
│   ├── duration_seconds         # 8, 16, or 24 (multiples of 8)
│   ├── generation_method        # "t2i_i2v" (characters) or "t2v" (environment only)
│   ├── video_prompt             # Veo 3, first 8s
│   ├── video_extend_prompt      # continuation if >8s
│   ├── t2i_prompt               # Nano Banana 2 reference image
│   ├── voiceover_text           # target language, humanized
│   ├── voiceover_duration_estimate_seconds
│   └── scene_description
└── cover_t2i_prompt         # thumbnail image prompt

Scene duration logic

• Veo 3 generates 8s max per clip
• 16s scene = 8s video looped 2x (ideal length)
• 24s scene = 8s video + extend prompt for continuation
• Ads, sponsors, and end cards are automatically filtered out

Generation methods

• t2i_i2v: Any scene with characters → generate reference image first (Nano Banana 2), then animate (Veo 3). Ensures visual consistency.
• t2v: Pure environmental/atmospheric scenes only → direct text-to-video.

Cost

Uses Gemini 2.5 Flash pricing. Token usage and USD cost are displayed after each run.

	Per 1M tokens
Input (video + text)	$0.30
Output	$2.50

Typical cost: $0.05 – $0.15 per analysis depending on video length.

Requirements

• Python 3.11+
• ffmpeg in PATH (for video merging)
• GEMINI_API_KEY environment variable

License

MIT