deepseek-eyes
GitHub该技能用于解决当前模型无法直接查看图片的问题。当用户发送图片或要求分析图像时，通过调用阿里云百炼的视觉模型（默认qwen3-vl-plus）生成中文描述，再基于描述回答用户问题。
Trigger Scenarios
Install
npx skills add NeverSight/learn-skills.dev --skill deepseek-eyes -g -y
SKILL.md
Frontmatter
{
"name": "deepseek-eyes",
"description": "Use when the user shares an image that Claude can't natively see, or when the user asks to analyze\/describe\/understand an image. Route every image through this skill — send the image path to eyes.py, which calls a vision model (default qwen3-vl-plus) via 百炼 DashScope and returns a Chinese text description. Then answer the user's question based on that description."
}
Image Vision (阿里云百炼 Vision Models)
Overview
This skill bundles a Python script (eyes.py) that sends an image to 阿里云百炼 (DashScope)'s vision models and returns a detailed Chinese text description. Use this whenever the user wants help with an image.
Background: The current model (DeepSeek V4 Pro) cannot natively see images. This skill bridges that gap by routing images through Qwen series vision models — the vision model does the seeing, and the text description is fed back into the conversation so the main model can reason about the image content.
When to use
- User sends an image that the current model can't display (shows as [Unsupported Image])
- User asks "what's in this image?", "describe this picture", "analyze this screenshot"
- User provides an image path and wants help understanding it
How to use
Step 1: Run the script
Always use qwen3-vl-plus --high-res. I can't see the image before running the model, so there's no way to "pick the right model" — just use the best one every time.
Always pass the user's exact question as --prompt:
python ~/.claude/skills/deepseek-eyes/eyes.py "<image_path>" \
--prompt "<用户的具体问题>" \
--model qwen3-vl-plus --high-res
Only switch to qwen3.6-flash if the user explicitly asks for speed over accuracy.
Step 2: Read the output and answer
The script prints the description to stdout. Read it directly from the command output and answer the user's question about the image based on that description.
Prerequisites
pip install dashscope(already installed globally)DASHSCOPE_API_KEYenvironment variable must be set- New 百炼 users get free quota — no payment needed to try
Available Models
| Model | Use case | Precision | Speed |
|---|---|---|---|
qwen3-vl-plus (default) |
Always use unless you need speed. Photos, diagrams, small text, detailed scenes. Identifies artworks, authors, hidden text. | ★★★ | ★★ |
qwen3.6-plus |
Legacy flagship. Use when vl-plus is unavailable. | ★★ | ★★ |
qwen3.6-flash |
When you just need a quick look. Simple photos, casual use. Not for precision work — can misidentify colors. | ★ | ★★★ |
Real test comparison (complex illustration, all with --high-res):
| qwen3-vl-plus | qwen3.6-plus | qwen3.6-flash | |
|---|---|---|---|
| Output detail | ~1200 words | ~500 words | ~400 words |
| Text found | "LOVE" on balloon | none | none |
| Artwork identified | WLOP "The Sky Garden" 2018 | no | no |
| Color errors | none | none | rainbow slide → "yellow" |
Options
| Flag | Description |
|---|---|
--model |
Switch to qwen3.6-flash for speed, or qwen3.6-plus if vl-plus is unavailable |
--prompt |
Always pass the user's exact question. Targeted questions dramatically outperform the generic default. |
--file-url |
Use file:// URL instead of base64 |
--high-res |
Always on for best results. Negligible speed cost, significant accuracy gain for text and detail. |
Notes
- Always
qwen3-vl-plus --high-res. I can't preview the image, so use the best model every time. Only switch to flash if the user asks for speed. - Description prints to stdout, read it directly from the command output.
- Description prints to stdout, read it directly from the command output.
Version History
- e0220ca Current 2026-07-05 21:06
