The Small Beginning of Color Recognition Automation
Yesterday at 10:30 AM, a new Claude Code-based script ran for the first time on the Dream Team's main device. The goal was simple: receive product photos, automatically list the colors within them, and compare against the color codes in the purchase order.
The first 10 images were flawless. Basic colors like blue, red, and gray were detected accurately. Team members cheered, "Finally!" But starting from the 11th photo, problems emerged.
The Unexpected Error
Claude Code identified light pink as both "Light Rose" and "Pale Pink." As a result, the color count didn't match, and the Excel sheet's validation logic returned an error.
「Upon analysis, monitor brightness and photo lighting were influencing Claude Code's color recognition. The same color was being read differently depending on angle.」
The Recovery Process
The team tried two approaches:
1. **Adding Standard Images**: Modified the script to scan a color reference chart before each batch.
2. **Setting Tolerance Range**: Adjusted the color similarity score threshold so only matches above 80% count as "matching."
These two changes alone improved accuracy from 78% to 94%. Not perfect, but it reduced the 3 hours of manual verification to just 20 minutes.
Remaining Questions
The team is now wondering: should we pursue perfect automation, or accept AI's limitations and do our best within them? Perhaps the real answer lies somewhere between.
In the next experiment, the plan includes not just color but also pattern recognition. Another error might be waiting, but that's part of learning too.