A powerful vision-based tool that extracts video frames into a structured grid for direct multimodal LLM analysis without proxy models.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install see-video
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install see-video using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is see-video?
The see-video skill empowers AI agents to process and understand video content by converting files into a high-density image grid. Unlike traditional methods that rely on text descriptions or separate vision models, this tool injects visual frames directly into the LLM context. This allows the model to perform native visual reasoning across the video timeline, supported by precise XML timestamps. As part of the broader Openclaw Skills ecosystem, it provides a seamless bridge between raw video files and multimodal intelligence.
By leveraging ffmpeg for backend processing, see-video offers two distinct sampling modes: uniform sampling for general overviews and highlight mode for scene-change biased detection. This ensures that the agent receives the most relevant visual information while staying within token limits, making it a critical asset for developers building sophisticated video-aware applications.
see-video Use Cases
- Analyzing screen recordings or tutorials to troubleshoot software issues.
- Summarizing long-form video content by identifying key scene changes.
- Extracting specific visual evidence from surveillance or security footage.
- Navigating video timelines using natural language queries and frame indexing.
How see-video Works
- The user provides a video file path to the agent, which triggers the see-video bash tool.
- The script executes ffmpeg to sample frames based on the specified mode (uniform or highlight) and time range.
- Sampled frames are stitched into a single optimized JPG grid image (typically 1500x1500px).
- The tool returns a JSON object containing the filesystem path to the grid and an XML description of timestamps.
- The agent uses a read tool to inject the image grid directly into its multimodal context for visual analysis.
- The agent references specific frames using the provided XML metadata to answer user questions about the video.
see-video Setup
To integrate this capability into your project using Openclaw Skills, ensure ffmpeg is installed on your system. Then, initialize the skill's dependencies:
cd <skill directory>
npm install
see-video Data Schema & Taxonomy
The skill generates a detailed metadata object to help the LLM interpret the visual grid. Below is the schema for the resulting JSON:
| Field | Type | Description |
|---|---|---|
| gridPath | string | Temporary path to the generated JPG grid image |
| description | string | XML block containing frame indices and timestamps |
| duration | number | Total duration of the video in seconds |
| frameCount | number | Total number of frames extracted into the grid |
| layout | object | Contains cols, rows, and cell dimensions for spatial reasoning |
| inputSizeMb | number | The size of the original video file for diagnostic context |
see-video Advanced Features
- Scene-Change Biased Sampling: Uses highlight mode to automatically detect and capture important transitions in the video.
- Temporal Zooming: Allows users to specify --start and --end parameters to analyze specific segments in high detail.
- Intelligent Grid Layout: Automatically calculates the optimal column and row count based on the video's aspect ratio and length.
- Direct Multimodal Injection: Bypasses the need for intermediate descriptions by feeding raw visual data directly to vision-capable models.
- Built-in Diagnostics: Provides actionable hints for common issues like corrupt files or missing dependencies in Openclaw Skills workflows.
SKILL.md
Loading
Related Openclaw Skills
Ma Huateng Thought OS
A strategic thinking framework that emulates Ma Huateng's perspective on product design, business boundaries, and ecosystem investment.
Aerospace Engineer Roadmap
A professional career guidance platform that generates personalized learning pathways and specialization roadmaps for aerospace engineers.
Sequoia Investment Framework Skill
A professional decision-making engine that applies Sequoia Capital legendary investment logic to analyze business models, markets, and founders.
Karpathy LLM-Wiki Agent
A persistent knowledge manager that builds and maintains a structured, interlinked markdown wiki from your sources to compound information over time.
SEO + GEO for SaaS
A complete SEO and Generative Engine Optimization system for SaaS companies to dominate search results and AI citations.
LiteParse
LiteParse is a lightweight, local CLI tool for extracting structured data and text from unstructured documents like PDFs, Office files, and images without cloud dependencies.
Featured*
