Qwen Vision Bridge

Claude cannot natively understand video. This skill bridges that gap by calling Qwen Omni — a natively multimodal model that processes video with temporal attention (it sees motion, not just individual frames).

The bridge also handles images, useful when you want Qwen's analysis on screenshots, diagrams, or photos.

How it works

A Python script at

${CLAUDE_PLUGIN_ROOT}/skills/qwen-vision/scripts/qwen_bridge.py

Prerequisites

The user must have:

1. DASHSCOPE_API_KEY

   environment variable set (get one at https://dashscope.console.aliyun.com/ or https://modelstudio.console.alibabacloud.com/)
2. Python 3.9+ with

   dashscope

   package installed

If the user hasn't set up yet, suggest running

/qwen-setup

Basic usage

python3 "${CLAUDE_PLUGIN_ROOT}/skills/qwen-vision/scripts/qwen_bridge.py" "/path/to/video.mp4" "Describe what happens in this video"

Parameters

--fps

--model

--json

--context

--save-context

--system-prompt

--prompt-file

Supported formats

Video: .mp4, .mov, .avi, .mkv, .webm, .flv, .wmv Image: .png, .jpg, .jpeg, .gif, .webp, .bmp, .tiff

Patterns

Single video analysis

python3 "${CLAUDE_PLUGIN_ROOT}/skills/qwen-vision/scripts/qwen_bridge.py" "/path/to/video.mp4" "Describe the character's body movement, poses, and transitions" --fps 2

Parse the text response and use it in your answer to the user.

Batch analysis

When the user has multiple videos to analyze, write a Python script that loops through files and calls the bridge for each one. Use

--json

references/batch-pattern.md

Multi-turn (follow-up questions)

# First question
python3 "${CLAUDE_PLUGIN_ROOT}/skills/qwen-vision/scripts/qwen_bridge.py" video.mp4 "General analysis" --save-context /tmp/ctx.json

# Follow-up
python3 "${CLAUDE_PLUGIN_ROOT}/skills/qwen-vision/scripts/qwen_bridge.py" video.mp4 "Tell me more about the lighting" --context /tmp/ctx.json

Image analysis

Same script, just pass an image path instead of video:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/qwen-vision/scripts/qwen_bridge.py" "/path/to/screenshot.png" "What UI elements are visible in this screenshot?"

Cost-saving tips

• Use

  --fps 1

  for long videos or when fine detail isn't needed
• Use

  --fps 0.5

  for very long videos (minutes+)
• For batch jobs, start with

  --fps 1

  and increase only if results are too vague

Error handling

• If

  DASHSCOPE_API_KEY

  is not set, the script exits with a clear error message. Guide the user to set it up.
• If

  dashscope

  is not installed, suggest

  pip install dashscope

  .
• If the API returns an error, the script prints the error code and message. Common issues: invalid key, quota exceeded, unsupported file format.
• If a video file is too large for the API, suggest lowering

  --fps

  or trimming the video first.

What Qwen sees vs what Claude sees

This is important context for the user: Qwen processes video frames with temporal attention — it understands motion, direction, rhythm, and transitions between frames. Claude analyzing individual screenshots cannot do this. When the user needs to understand what happens in a video (not just what a single frame looks like), this bridge is the right tool.

Additional resources

• references/batch-pattern.md

  — template for batch video classification

• references/prompt-tips.md

  — effective prompts for different analysis types