Crop Tool

Overview

• Analyze charts and read small text/values
• Extract data from documents with fine details
• Compare specific regions in complex images
• Examine technical diagrams closely
• Process images with multiple focal points

What It Does

1. Request crops — Claude specifies a region using normalized coordinates (0-1)
2. Execute crops — Tool extracts the specified region from the image
3. Analyze details — Claude examines the cropped region in detail
4. Iterate — Claude crops additional regions as needed for complete analysis

Coordinate System

• (0, 0) = top-left corner
• (1, 1) = bottom-right corner
• (0.5, 0.5) = center of image

How to Use This Skill

Basic Usage

"Analyze this chart and tell me which bar is tallest."

1. View the full image
2. Identify which regions need detail
3. Crop those regions
4. Examine cropped images
5. Provide detailed analysis

Scenario 1: Reading Small Text

1. Crop the legend region
2. Read text in high detail
3. Return the values

Scenario 2: Comparing Values

1. Examine full pie chart
2. Crop uncertain regions
3. Compare with precision
4. Provide answer with confidence

Scenario 3: Dense Document Analysis

1. Examine full document layout
2. Crop tables and data sections
3. Extract precise values
4. Organize results

Technical Details

Tool Specification

{
  "name": "crop_image",
  "description": "Crop an image by specifying a bounding box",
  "input_schema": {
    "type": "object",
    "properties": {
      "x1": {
        "type": "number",
        "minimum": 0,
        "maximum": 1,
        "description": "Left edge (0 = left, 0.5 = center, 1 = right)"
      },
      "y1": {
        "type": "number",
        "minimum": 0,
        "maximum": 1,
        "description": "Top edge (0 = top, 0.5 = center, 1 = bottom)"
      },
      "x2": {
        "type": "number",
        "minimum": 0,
        "maximum": 1,
        "description": "Right edge of bounding box"
      },
      "y2": {
        "type": "number",
        "minimum": 0,
        "maximum": 1,
        "description": "Bottom edge of bounding box"
      }
    },
    "required": ["x1", "y1", "x2", "y2"]
  }
}

Requirements

• Image format: PNG, JPEG, GIF, WebP
• Claude model: Opus 4.5+ recommended (supports improved vision)
• Dependencies: PIL/Pillow for image processing

How It Works

1. Normalization → Convert normalized coordinates (0-1) to pixels
2. Extraction → Crop specified region from image
3. Encoding → Convert cropped image to base64
4. Return → Send cropped image back to Claude
5. Iteration → Claude can crop multiple regions

Performance Impact

• ✅ Better accuracy on small text (95%+ vs 70%)
• ✅ Faster analysis of complex images
• ✅ More reliable chart/data extraction
• ✅ Improved handling of dense documents

• Charts with legend/axis labels
• Documents with tables
• Technical diagrams
• Images with multiple focal points
• Any task requiring precision on image details

Examples

Example 1: Chart Analysis

Input: Pie chart with small color legend
Question: "Which color represents the minimum value?"

Process:
1. Claude views full chart
2. Crops legend region (0.8-1.0, 0.0-0.4)
3. Reads color-value mappings
4. Analyzes main chart
5. Returns answer with confidence

Example 2: Document Extraction

Input: Technical document with multiple tables
Task: "Extract all values from the comparison table"

Process:
1. Claude scans document layout
2. Crops each table region
3. Reads values precisely
4. Returns structured data

Example 3: Multi-region Analysis

Input: Complex dashboard with multiple charts
Task: "Compare trends across all sections"

Process:
1. Claude views full dashboard
2. Crops first chart region
3. Analyzes first chart
4. Crops second chart region
5. Analyzes second chart
6. Compares across regions
7. Returns synthesis

When to Use vs Not Use

Use crop tool when:

• ✅ Image contains small text or fine details
• ✅ Need high precision on specific regions
• ✅ Analyzing charts, tables, or documents
• ✅ Multiple focal points in image
• ✅ Data extraction required

Not needed when:

• ❌ Full image already in focus
• ❌ Large text is readable at full resolution
• ❌ Simple object recognition tasks
• ❌ General scene understanding

Architecture

Image Input
    ↓
Claude Views Full Image
    ↓
Claude Decides Regions to Crop
    ↓
Tool Executes Crop (normalized → pixels)
    ↓
Cropped Image Encoded as Base64
    ↓
Claude Analyzes Crop
    ↓
Iterate Until Complete
    ↓
Final Analysis

Tips for Best Results

1. Provide clear context — Tell Claude what you're analyzing
2. Mention details needed — Highlight if precision matters
3. Ask for reasoning — Request Claude show cropped regions
4. Verify accuracy — Check Claude's extracted values
5. Chain crops — Let Claude crop iteratively as needed

Integration with Other Tools

• Document tools (pdf, docx, xlsx) — Extract data from complex layouts
• Web artifacts — Analyze UI screenshots
• Code review — Examine code snippets in images
• Any vision task — Improve accuracy through detail

Feedback & Improvements