The Problem: The Nightmare of Label Image Management
A small non-developer team faced a repetitive daily task: manually entering accumulated shipping label images into a spreadsheet every week.
• 200 images × 3 data points (tracking number, destination, quantity) = roughly 600 keystrokes
• Repeated twice weekly
• Monthly time loss: approximately 4,800 minutes (80 hours)
One day someone asked, "Can't Claude Code read images too?"
First Attempt: Automated Image Upload
Using Claude Code's file processing capabilities, the team built this workflow:
1. Collect images from local Windows device folder
2. Batch send images to Claude Code
3. Extract recognized text as JSON format
4. Auto-insert into Google Sheets
Result? The first 50 images came out perfectly accurate.
Unexpected Errors Discovered
1. Number Confusion in Low-Resolution Images (O vs. 0)
A tracking number was read as `1234O567` when it was actually `12340567`.
Solution: Added pre-processing logic before image transmission with a minimum 300 DPI filter standard.
2. Text Recognition Failure on Tilted Angles
Shipping addresses were only 50% extracted when labels weren't perfectly horizontal to the camera.
Improvement: Added to the Claude Code prompt: "If text appears tilted, please correct it as much as possible."
3. Duplicate Processing (Same File Handled Twice)
Early automation lacked completion logs, causing identical tracking numbers to appear 2-3 times in the spreadsheet.
The Solution
Automatically move processed images to a `_processed` folder and record a "processing timestamp" in column B of the spreadsheet.
Final Process
3 Steps Even Non-Developers Can Follow
Step 1: Prepare Images
• Collect shipping label photos in one folder
• Name files with numbers for easy sorting (example: `001_shipping.jpg`)
Step 2: Write Your Claude Code Prompt
Start with instructions like this:
"Please read all JPG files in this folder. Extract the following from each image and output as JSON:
• Tracking number (including hyphens)
• Destination country
• Quantity (numbers only)
For anything uncertain, please mark it [uncertain]."
Step 3: Verify Results and Insert into Sheets
After Claude Code runs, connect the JSON output via Google Sheets' Apps Script for auto-insertion.
Verification is essential: only manually check lines marked "[uncertain]" (typically 5-10 rows).
One Month of Results
• Processing time: 80 hours → 8 hours (90% reduction)
• Accuracy rate: 96% (4% manual corrections)
• Team feedback: "Finally we have time to actually respond to buyers"
Key Takeaway
Automation isn't about achieving perfection, it's about reaching "good enough." At 96% accuracy, you only need to manually check 4%, making the overall speed vastly superior to fully manual work.
Claude Code's image reading capability proved more powerful than expected. The key was being explicit in the prompt: "Be honest when uncertain" significantly improved transparency.
Next experiment: multi-language label recognition. Can it identify and categorize labels mixing Traditional Chinese, English, and Simplified Chinese?