What Happened with 10K Rows
While running Google Sheets automation on a Windows device, we encountered an unexpected problem. When processing 10,000 rows of monthly transaction records with Claude Code for sorting and validation, the team flagged that certain price ranges were being missed.
"Claude Code caught last month's price range fine, so why can't it find this month's specific bracket?"
Root Cause: Conditional Logic Gap
The issue was simpler than expected. The prompt given to Claude Code was:
"Find data that is 300 or above and below 500"
But the actual data contained the exact value "300," and when Claude Code interpreted "300 or above," it introduced subtle floating-point precision errors during calculation. A small gap formed between the non-developer's expected logic and what AI actually executed.
Solutions: Two Experiments
First Attempt (Failed)
We rewrote the prompt with greater clarity.
Check the price column in Google Sheets,
Highlight all data where 300 <= price < 500.
Eliminate unnecessary decimal-point calculations.
Result: Only about 70% of data detected. Claude Code understood the conditions, but when processing 10K rows at once, the workflow seemed to break mid-process.
Second Attempt (Success)
We decided to break the large dataset into smaller chunks.
1. First, divide Google Sheets into 5 sections of 1,000 rows each
2. Run Claude Code on each section separately
3. Collect results in separate tabs
4. Perform final integration verification
With this approach, "missed" data nearly disappeared. Processing just 1,000 rows at a time raised Claude Code's accuracy above 95%.
Key Takeaways
Two critical realizations from a non-developer's perspective:
1. **Claude Code is a "quick assistant," not "perfect automation."** Especially when handling tens of thousands of rows, skepticism like "should we split this?" is actually smarter than betting on "it'll work first try."
2. **Adjusting work unit size matters as much as prompt clarity.** Simply reducing data scope per task dramatically improves accuracy.
This experience reinforced that automation tools don't fully replace non-developer work, they require balancing human judgment with tool capability.