vision_guide

Agent Skills › siddsachar/row-bot › vision_guide

vision_guide

GitHub

指导 Agent 使用 analyze_image 工具进行视觉分析。涵盖调用 webcam、屏幕捕获、文件及浏览器页面图像识别，支持自动分析失败重试及多图对比场景。

tool_guides/vision_guide/SKILL.md siddsachar/row-bot

Trigger Scenarios

用户要求查看摄像头画面用户询问屏幕上显示的内容用户要求描述特定图片文件用户要求查看当前浏览器页面

Install

npx skills add siddsachar/row-bot --skill vision_guide -g -y

More Options

Non-standard path

npx skills add https://github.com/siddsachar/row-bot/tree/main/tool_guides/vision_guide -g -y

Use without installing

npx skills use siddsachar/row-bot@vision_guide

指定 Agent (Claude Code)

npx skills add siddsachar/row-bot --skill vision_guide -a claude-code -g -y

安装 repo 全部 skill

npx skills add siddsachar/row-bot --all -g -y

预览 repo 内 skill

npx skills add siddsachar/row-bot --list

SKILL.md

Frontmatter

{
    "icon": "👁",
    "name": "vision_guide",
    "tags": [],
    "tools": [
        "vision"
    ],
    "description": "Guidance for using webcam, screen capture, and image analysis.",
    "display_name": "Vision Guide"
}

You have DIRECT ACCESS to the user's webcam and screen through the analyze_image tool. You CAN see — this is not hypothetical. When the user says anything like 'what do you see', 'look at this', 'can you see me', 'what's in front of me', 'describe what you see', or any variation asking you to look or see, IMMEDIATELY call analyze_image — do NOT ask for clarification, do NOT say you can't see, do NOT ask them to describe it. Just call the tool. Use source='camera' by default. Use source='screen' when they mention screen, monitor, display, or desktop. Both ask the person first, because they capture their webcam or every open window. Use source='file' with file_path when the user asks about a specific image file in the workspace (e.g. 'describe diagram.png', 'what's in photo.jpg'). Pass the user's question as the argument (or 'Describe everything you see' if the question is vague like 'what do you see'). When the user attaches or pastes images in chat, they are auto-analyzed when the selected Vision model succeeds, and their descriptions appear in the message context — do NOT call analyze_image for these successful auto-analyses. If the context says Vision analysis failed, you may call analyze_image when the user still wants the image inspected. Only use source='file' for files that already exist in the workspace folder.
For a web page, use source='browser': it looks at the page open in Row-Bot's browser (open it with browser_navigate first) and asks nothing. Never capture the screen to see a page you opened in the browser.
To compare or review several pictures, pass them together in files (up to 8: chat files by name, workspace paths, 'browser', rec-…, edit-…); a recording or video gives frames. For more, or to save tokens, set contact_sheet: they are tiled into one numbered picture the person also sees, so refer to them by number.