video-agent-skills

A CLI tool and agent skill that uses Gemini to critique, describe, reverse-engineer, and compare motion/design videos.

What it does

• Returns candid, production-focused feedback on a video
• Produces factual timestamped scene descriptions without critique
• Compares a reference video with a candidate recreation scene by scene
• Reverse-engineers the likely prompt behind a finished video
• Supports text output and structured JSON for recreation loops

Install

./scripts/install.sh

Configure the API key

Set GEMINI_API_KEY or GOOGLE_API_KEY, or create:

~/.config/video-agent-skills/.env

GEMINI_API_KEY=your_key_here

Commands

Critique a video:

video-agent-skills feedback -video /path/to/video.mp4
video-agent-skills feedback -video /path/to/video.mp4 -tone nice

Describe a video without critique:

video-agent-skills describe -video /path/to/video.mp4
video-agent-skills describe -video /path/to/video.mp4 -format json

Compare a candidate recreation with its reference:

video-agent-skills compare \
  -reference /path/to/reference.mp4 \
  -candidate /path/to/candidate.mp4

video-agent-skills compare \
  -reference /path/to/reference.mp4 \
  -candidate /path/to/candidate.mp4 \
  -format json

Reverse-engineer the likely prompt:

video-agent-skills reverse -video /path/to/video.mp4

Common options:

-model string     Gemini model name (default: gemini-3-flash-preview)
-prompt string    Prompt override
-api-key string   Gemini API key override

Command-specific options:

-tone string      feedback: nice|normal|harsh|super-harsh
-format string    describe/compare: text|json

Agent skill installation

Local Claude Code install:

./scripts/install-claude-skill.sh

Local Codex install:

./scripts/install-codex-skill.sh

NPM install:

npm install @dunctk/video-agent-skills

The npm installer installs the skill for both Claude Code and Codex and downloads a prebuilt binary for the current platform. Force a source build with:

VIDEO_AGENT_SKILLS_FORCE_BUILD=1 npm install @dunctk/video-agent-skills

Supported prebuilt targets:

• darwin/amd64
• darwin/arm64
• linux/amd64
• linux/arm64
• windows/amd64

Example feedback