vision（vision 图像）

Vision Image Processing Skill

Overview

This skill leverages Claude's multimodal vision capabilities to analyze, process, and extract insights from images. It supports a wide range of visual understanding tasks including optical character recognition (OCR), image classification, diagram analysis, chart interpretation, and visual comparison.

When to Use This Skill

Activate this skill when users need to:

• Extract text from images, screenshots, or scanned documents (OCR)
• Classify or categorize images based on visual content
• Analyze charts, graphs, or data visualizations to extract insights
• Compare multiple images (diagrams, screenshots, designs)
• Describe or caption images in detail
• Answer questions about visual content
• Detect objects, people, or elements within images
• Analyze UI/UX from screenshots or mockups
• Read handwritten text or notes
• Process receipts, invoices, or forms for data extraction

Core Capabilities

1. Optical Character Recognition (OCR)

Extract text from images with high accuracy:

Instructions:

• Use the Read tool to load the image file
• Analyze the image and extract all visible text
• Preserve formatting, layout, and structure when possible
• Handle multiple languages and fonts
• Identify and extract text from challenging contexts (handwriting, artistic fonts, rotated text)

Output Format:

• Provide extracted text in markdown format
• Include confidence notes for challenging sections
• Maintain document structure (headings, paragraphs, lists)

Example Use Cases:

• Screenshot text extraction
• Scanned document digitization
• Receipt and invoice processing
• Handwritten note transcription
• Sign and label reading

2. Image Classification and Categorization

Identify and classify image content:

Instructions:

• Analyze the overall subject and context
• Identify primary objects, scenes, or themes
• Provide classification labels with confidence levels
• Detect style, mood, and artistic elements
• Categorize by industry-relevant taxonomies when applicable

Output Format:

## Primary Classification
- Category: [main category]
- Confidence: [High/Medium/Low]

## Detected Elements
- Object 1: [description]
- Object 2: [description]
...

## Additional Attributes
- Style: [style description]
- Setting: [environment/context]
- Colors: [dominant colors]

3. Chart and Graph Analysis

Extract insights from data visualizations:

Instructions:

• Identify chart type (bar, line, pie, scatter, etc.)
• Extract data points, values, and trends
• Read axes labels, legends, and annotations
• Summarize key insights and patterns
• Flag anomalies or notable data points

Output Format:

## Chart Analysis

**Type:** [Chart Type]

**Data Summary:**
[Extracted data in table or structured format]

**Key Insights:**
1. [Insight 1]
2. [Insight 2]
3. [Insight 3]

**Trends:**
- [Trend description]

**Notable Points:**
- [Anomalies or important observations]

4. Diagram and Visual Comparison

Compare multiple images or diagrams:

Instructions:

• Load all images to be compared
• Identify similarities and differences
• Highlight structural, content, and style variations
• Create side-by-side comparison tables
• Note additions, deletions, or modifications

Output Format:

## Visual Comparison

**Image 1:** [description]
**Image 2:** [description]

### Similarities
- [Similarity 1]
- [Similarity 2]

### Differences
| Aspect | Image 1 | Image 2 |
|--------|---------|---------|
| [Aspect] | [Description] | [Description] |

### Overall Assessment
[Summary of comparison]

5. Detailed Image Description

Generate comprehensive image descriptions:

Instructions:

• Describe the overall scene or subject
• Identify and describe all visible elements
• Note spatial relationships and composition
• Describe colors, lighting, and atmosphere
• Mention text, logos, or symbols if present
• Consider accessibility (generate alt-text compatible descriptions)

Output Format:

• Natural language description (paragraph form)
• Structured element list (bulleted)
• Technical details (dimensions, format, quality notes)

6. Visual Question Answering

Answer specific questions about image content:

Instructions:

• Carefully read the user's question
• Examine the relevant areas of the image
• Provide accurate, specific answers
• Reference visual evidence when answering
• Acknowledge uncertainty if details are unclear

Best Practices:

• Be precise and factual
• Avoid assumptions beyond what's visible
• Describe what you see, not what you infer (unless asked)
• Use spatial language (top-left, center, background, etc.)

7. UI/UX and Design Analysis

Analyze user interfaces and design elements:

Instructions:

• Identify UI components (buttons, forms, navigation)
• Assess layout and visual hierarchy
• Note design patterns and conventions
• Evaluate accessibility considerations
• Compare against design best practices
• Extract color schemes and typography

Output Format:

## UI/UX Analysis

**Component Inventory:**
- [List of UI elements]

**Layout Assessment:**
- [Layout description and grid analysis]

**Design Patterns:**
- [Identified patterns]

**Accessibility Notes:**
- [Contrast, readability, touch targets]

**Recommendations:**
- [Improvement suggestions]

8. Document and Form Processing

Extract structured data from forms, receipts, and documents:

Instructions:

• Identify document type and structure
• Extract field names and values
• Organize data into structured format (JSON, CSV, tables)
• Handle multi-column layouts
• Preserve data relationships and hierarchies

Output Format:

{
  "document_type": "invoice",
  "fields": {
    "invoice_number": "value",
    "date": "value",
    "total": "value"
  },
  "line_items": [...]
}

Workflow and Best Practices

Standard Vision Processing Workflow

1. Load the Image(s)

- Use the Read tool to access image files - Support formats: PNG, JPG, JPEG, GIF, WebP, PDF (single page)

1. Understand the Request

- Identify the specific task (OCR, classification, analysis, etc.) - Note any special requirements or focus areas

1. Analyze the Visual Content

- Apply Claude's vision capabilities to examine the image - Extract relevant information based on the task

1. Structure the Output

- Format results according to the task type - Use markdown for readability - Include confidence indicators where appropriate

1. Validate and Refine

- Check for completeness - Verify accuracy of extracted data - Provide follow-up options if needed

Quality Guidelines

• Accuracy First: Prioritize correct information over comprehensive coverage
• Structured Output: Use consistent formatting for similar tasks
• Confidence Indicators: Note when details are unclear or ambiguous
• Context Awareness: Consider the user's domain and use case
• Accessibility: Generate descriptions suitable for screen readers when appropriate

Limitations and Considerations

• Image Quality: Low resolution or blurry images may reduce accuracy
• Supported Formats: Primarily raster images; vector graphics may need conversion
• Privacy: Be cautious with sensitive information (PII, credentials, etc.)
• Complex Diagrams: Highly technical diagrams may require domain expertise clarification
• Real-Time Data: Cannot access live data or external resources not in the image

Advanced Features

Batch Processing

For multiple images:

Processing images in batch:
1. [Image1.png] - [Task result]
2. [Image2.png] - [Task result]
3. [Image3.png] - [Task result]

Summary: [Overall findings]

Multi-Modal Context

Combine visual analysis with code, documents, or data:

• Cross-reference image content with codebase files
• Validate design implementations against mockups
• Extract data and populate code templates
• Generate code from UI screenshots

Helper Script Integration

For advanced processing, the skill includes a Python helper script at scripts/image_processor.py:

• Image format conversion
• Metadata extraction
• Batch file operations
• Image preprocessing (resize, enhance, crop)

Usage:

python .claude/skills/vision/scripts/image_processor.py --operation [convert|metadata|batch] --input [path]

Examples

Example 1: OCR from Screenshot

User Request: "Extract the text from this screenshot"

Process:

1. Load image with Read tool
2. Identify all text regions
3. Extract and format text
4. Preserve structure

Output:

## Extracted Text

### Heading
Main paragraph text goes here...

### List Items
- Item 1
- Item 2
- Item 3

---
Note: Text extracted with high confidence. Minor formatting adjustments made for readability.

Example 2: Chart Analysis

User Request: "Analyze this sales chart and tell me the key trends"

Process:

1. Load chart image
2. Identify chart type and data
3. Extract values and labels
4. Analyze trends

Output:

## Sales Chart Analysis

**Chart Type:** Line graph with 3 series (2023-2025)

**Data Points:**
| Month | 2023 | 2024 | 2025 |
|-------|------|------|------|
| Jan   | 120K | 145K | 180K |
| Feb   | 135K | 150K | 195K |
...

**Key Insights:**
1. Consistent year-over-year growth of ~20%
2. Seasonal peak in Q4 across all years
3. 2025 showing accelerated growth trend

**Recommendation:** Prepare for continued high demand in Q4 2025

Example 3: UI Component Comparison

User Request: "Compare these two design mockups and highlight the differences"

Process:

1. Load both images
2. Identify UI components in each
3. Compare layouts and elements
4. Document changes

Output:

## Design Mockup Comparison

### Added Elements (Mockup 2)
- Search bar in header
- User avatar in top-right corner
- Breadcrumb navigation

### Removed Elements
- Sidebar menu (replaced with top navigation)

### Modified Elements
| Component | Mockup 1 | Mockup 2 |
|-----------|----------|----------|
| Primary Button | Blue (#007AFF) | Green (#34C759) |
| Typography | Roboto 16px | SF Pro 15px |
| Card Spacing | 16px | 24px |

### Layout Changes
- Shifted from sidebar to top navigation
- Increased content area width by ~30%
- Changed from 3-column to 4-column grid

Integration with Claude Code

This skill works seamlessly with other Claude Code features:

• Read Tool: Load images from the filesystem
• Write Tool: Save processed results or extracted data
• Bash Tool: Run helper scripts for preprocessing
• Task Tool: Coordinate complex multi-image workflows

Quick Reference

Tips for Best Results

1. Provide Context: Mention the domain or purpose (e.g., "medical diagram," "e-commerce UI")
2. Be Specific: Request specific information rather than general analysis
3. Multiple Angles: For complex images, ask follow-up questions
4. File Paths: Use absolute or relative paths correctly
5. Batch Operations: Process multiple similar images together for consistency

Support and Troubleshooting

Common Issues:

• "Cannot read image" → Verify file path and format
• "Low confidence extraction" → Image may be too low resolution
• "Unable to detect chart data" → Chart may be too complex or stylized

Getting Better Results:

• Use high-resolution images (300+ DPI for documents)
• Ensure good contrast and lighting
• Crop images to focus on relevant areas
• Provide context about the image content

License

This skill is licensed under Apache-2.0.

Version

Version: 1.0.0 Last Updated: 2025-11-18 Compatible with: Claude Code (all versions with vision support)