watch

name	watch
description	Use when asked to analyze, review, or describe a local video file (e.g., .mp4, .mov). Describes how to use the `mt watch` command to stream large videos to Gemini securely with server-side context caching and multi-turn thread support.

Video Analysis via mt watch

Overview

Use this skill when you need to analyze a local video file. Large video files cannot be processed directly in your local context or uploaded as standard attachments without hitting token and size limits. The mt watch command provides a dedicated workflow to securely upload the video in chunks, cache it server-side, and stream Gemini model responses.

Workflow

1) Initial Analysis

To start a new analysis on a video, run mt watch with the file path and your prompt:

mt watch "/path/to/video.mp4" "Describe what happens in this video in detail."

If the prompt is long or complex, you can construct it first and pass it via a variable:

PROMPT="Look closely at the UI transitions. What color is the primary button before and after the click?"
mt watch "/path/to/video.mp4" "$PROMPT"

The command will output the model's text response. Note: For machine-readable output or extra diagnostics (like file deduplication or cache telemetry), append --json or --verbose.

2) Follow-up Questions (Multi-turn)

When you run mt watch with the --verbose flag (or inspect the JSON output), you will receive a thread_id. You can use this thread_id to ask follow-up questions without re-uploading or re-processing the entire video from scratch.

# Add --verbose to see the thread_id
mt watch "/path/to/video.mp4" "Describe the video" --verbose

# Use the returned thread_id for a follow-up:
mt watch "/path/to/video.mp4" "What was the text on the second screen?" --thread <thread_id>

Tips & Constraints

• Supported Formats: .mp4, .mov, .webm
• Deduplication: The server hashes the file. If it has been uploaded recently, it skips the upload phase and jumps straight to analysis.
• Context Limits: If the video is extremely long, ask specific, targeted questions rather than broad "describe everything" prompts to get the best results from the model.