Image support

Overview

• Claude can process images included in user messages
• Possible tasks: describe image content, compare images, count objects, and more

Technical Limitations & Requirements

• Maximum 100 images per single request
• Each image has size and dimension restrictions
• Images consume tokens; token count can be calculated using an equation based on image height and width in pixels

Sending Images to Claude

• Images are sent using image blocks within user messages
• Multiple image blocks can be included in one message
• Each image block contains either:
• Raw image data (base64 encoded)
• URL to an online-hosted image

Prompting Techniques for Images

• Simple prompts with images often produce poor results
• Strong prompting techniques dramatically improve accuracy, including:
• Providing analysis steps for Claude to follow
• Using one-shot or multi-shot examples
• Giving clear guidelines

Example: Step-by-Step Analysis

• Break down the task into sequential steps (e. g. , identify each object individually, recount using a different method, compare results)
• Tested example: improved marble counting from incorrect count of 13 to correct count of 12

Example: One-Shot/Multi-Shot Prompting

• Alternate image blocks and text blocks in the same message
• Provide example image with correct answer before asking Claude to analyze the target image
• Improves accuracy on subsequent image analysis

Practical Use Case: Fire Risk Assessment

• Automate property inspections using satellite imagery instead of in-person visits
• Claude can analyze: main residence location, tree density, fire service access, overhanging branches, and overall fire risk
• Assign risk ratings (1-4) based on multiple criteria
• Provide structured output with summary and score

Key Takeaway

• Success with images depends on prompt quality, not just image input
• Apply the same prompt engineering principles used for text-based tasks

Takeaways

• Claude can process up to 100 images per request using image blocks containing base64-encoded data or URLs
• Strong prompting techniques dramatically improve image analysis accuracy—use step-by-step instructions, examples, or multi-shot prompting rather than simple prompts
• Images consume tokens based on their dimensions; token count can be calculated using a formula based on pixel height and width
• Claude can automate practical tasks like fire risk assessment from satellite imagery by analyzing multiple criteria and assigning structured risk ratings
• Success with images depends on prompt quality and engineering, not just image input alone