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 sends media files to the Qwen API and returns the analysis as text. Call it via Bash.

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 first.

Basic usage

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

Parameters

Flag	Default	Description
(positional 1)	required	Path to video or image file
(positional 2)	generic prompt	Analysis prompt
--fps	2.0	Frames per second to sample from video. Lower = cheaper, higher = more detail
--model	qwen-omni-plus-latest	Qwen model to use
--json	off	Output as JSON (for parsing)
--context	none	Path to JSON file with previous conversation (multi-turn)
--save-context	none	Save conversation context for follow-up questions
--system-prompt	none	Custom system prompt for Qwen
--prompt-file	none	Read prompt from a file instead of argument

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 flag for machine-readable output. See references/batch-pattern.md for a template.

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