vision

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Vision (delegated image understanding)

Vision（委托式图片理解）

You do not have native vision, but you can still "see" an image by running the bundled script, which sends the image to a configurable OpenAI-compatible vision model and returns a text answer.

您不具备原生视觉能力，但可以通过运行附带的脚本“看见”图片，该脚本会将图片发送至可配置的兼容OpenAI的视觉模型，并返回文本结果。

When to use

使用场景

Use this skill whenever a task requires understanding image content and you cannot view it directly, for example:

The user uploads or points to an image / screenshot / photo and asks what's in it.
You need to read text inside an image (OCR).
You need to diagnose an error from a screenshot.
You need to understand a UI mockup, diagram, chart, or scanned document.
You need to compare what an image shows against code or expected output.

当任务需要理解图片内容而您无法直接查看时，即可使用此技能，例如：

用户上传或指向某张图片/截图/照片并询问其中内容。
您需要提取图片中的文字（OCR）。
您需要通过截图诊断错误。
您需要理解UI原型、示意图、图表或扫描文档。
您需要对比图片显示内容与代码或预期输出。

How to use

使用方法

Run the script with the Bash tool. Pass the image (local path or http(s) URL) and a clear, specific instruction describing what you need to know:

bash

python3 "$CLAUDE_SKILL_DIR/scripts/see.py" <image_path_or_url> "your question"

$CLAUDE_SKILL_DIR

is not set in your environment, use the relative path to this skill folder, e.g.

python3 scripts/see.py ...

from the skill directory, or the absolute path where the skill is installed.

Examples:

bash

undefined

使用Bash工具运行脚本。传入图片（本地路径或http(s) URL）以及清晰、明确的指令说明您需要了解的信息：

bash

python3 "$CLAUDE_SKILL_DIR/scripts/see.py" <image_path_or_url> "your question"

如果您的环境中未设置

$CLAUDE_SKILL_DIR

，请使用此技能文件夹的相对路径，例如在技能目录下使用

python3 scripts/see.py ...

，或使用技能安装的绝对路径。

示例：

bash

undefined

Describe an image

描述图片内容

python3 scripts/see.py ./photo.jpg "Describe this image in detail"

OCR — extract text

OCR——提取文字

python3 scripts/see.py ./receipt.png "Transcribe all text exactly, preserving layout"

Diagnose an error screenshot

诊断错误截图

python3 scripts/see.py ./error.png "What error is shown and what is the likely cause?"

Read a chart into structured data

将图表内容转换为结构化数据

python3 scripts/see.py ./chart.png "Extract every series, label, and value as a markdown table"

Remote image

远程图片

python3 scripts/see.py "https://example.com/diagram.png" "Explain this architecture diagram"


The script prints the model's answer to stdout. Read that answer and use it to
continue the task. Ask a focused question rather than a generic "describe" when
you need something specific (a value, a status, an error message) — you get
better results and spend fewer tokens.

python3 scripts/see.py "https://example.com/diagram.png" "Explain this architecture diagram"


脚本会将模型的结果打印到标准输出。读取该结果并继续完成任务。当您需要特定信息（某个数值、状态、错误信息）时，请提出针对性问题，而非泛泛的“描述”——这样能获得更好的结果，且消耗更少的token。

Configuration (required, set once)

配置（必填，仅需设置一次）

The script reads its config from environment variables or a

.claude/settings.json

env

block (same variable names as the vision-mcp-server MCP, so config carries over). Resolution order: an explicit environment variable wins; otherwise the script reads

.claude/settings.json

, searching from the current directory upward and then

~/.claude/

Variable	Required	Example
`VISION_BASE_URL`	yes	`http://localhost:1234/v1/chat/completions`
`VISION_MODEL`	yes	`Qwen3-VL-32B` , `gpt-4o` , `glm-4v` , ...
`VISION_API_KEY`	no*	your API key (*optional for local servers)
`VISION_MAX_TOKENS`	no	`4096`
`VISION_TEMPERATURE`	no	`0.2`
`VISION_DETAIL`	no	`auto` \| `low` \| `high`
`VISION_TIMEOUT`	no	`120`

VISION_BASE_URL
must be the full chat-completions endpoint (
.../v1/chat/completions
), not just the base URL.

Set them in your shell profile, or in the MCP/agent

env

block, or inline:

bash

export VISION_BASE_URL=http://localhost:1234/v1/chat/completions
export VISION_MODEL=Qwen3-VL-32B
export VISION_API_KEY=sk-...        # optional for local

Or put them in

.claude/settings.json

(project-level, or global

~/.claude/

jsonc

{
  "env": {
    "VISION_BASE_URL": "http://localhost:1234/v1/chat/completions",
    "VISION_MODEL": "Qwen3-VL-32B",
    "VISION_API_KEY": "sk-..."
  }
}

脚本从环境变量或

.claude/settings.json

的

env

块中读取配置（变量名称与vision-mcp-server MCP相同，因此配置可通用）。优先级：显式设置的环境变量优先；否则脚本会从当前目录向上搜索

.claude/settings.json

，若未找到则查找

~/.claude/

目录下的文件。

变量名	是否必填	示例
`VISION_BASE_URL`	是	`http://localhost:1234/v1/chat/completions`
`VISION_MODEL`	是	`Qwen3-VL-32B` , `gpt-4o` , `glm-4v` , ...
`VISION_API_KEY`	否*	您的API密钥（*本地服务器可选）
`VISION_MAX_TOKENS`	否	`4096`
`VISION_TEMPERATURE`	否	`0.2`
`VISION_DETAIL`	否	`auto` \| `low` \| `high`
`VISION_TIMEOUT`	否	`120`

VISION_BASE_URL
必须是完整的聊天补全端点（
.../v1/chat/completions
），而非仅基础URL。

您可以在shell配置文件中设置这些变量，或在MCP/agent的

env

块中设置，也可直接在命令行中设置：

bash

export VISION_BASE_URL=http://localhost:1234/v1/chat/completions
export VISION_MODEL=Qwen3-VL-32B
export VISION_API_KEY=sk-...        # 本地服务器可选

或者将其放入

.claude/settings.json

中（项目级或全局

~/.claude/

目录下）：

jsonc

{
  "env": {
    "VISION_BASE_URL": "http://localhost:1234/v1/chat/completions",
    "VISION_MODEL": "Qwen3-VL-32B",
    "VISION_API_KEY": "sk-..."
  }
}

Notes

注意事项

Pure Python standard library — no
```
pip install
```
needed.
Local files are auto-converted to a base64 data URL; http(s) URLs are passed through.
If the script reports a missing variable or an unreachable endpoint, fix the config above and retry. Add
```
--dry-run
```
to inspect the request without sending it:
```
python3 scripts/see.py --dry-run img.png "test"
```
.

基于纯Python标准库——无需执行
```
pip install
```
安装依赖。
本地文件会自动转换为base64数据URL；http(s)URL会直接传递。
如果脚本提示缺少变量或端点不可达，请修复上述配置并重试。添加
```
--dry-run
```
参数可在不发送请求的情况下检查请求内容：
```
python3 scripts/see.py --dry-run img.png "test"
```
。