That Day's Small Incident
One of our automation agents suddenly started failing to process orders. The same data would intermittently throw errors. At first, we thought it was a Google Sheets API connection issue, but the cause was completely different.
On a Windows device's Google Sheets, we had hidden certain columns, but Claude Code wasn't detecting the existence of these hidden columns.
Why This Happened
Google Sheets API does return hidden column data. But when Claude Code reads the sheet, it either skips or misinterprets hidden sections. In the end, the data structure we designed didn't match what the agent received.
[Visible Data]
Column A(Customer Name) → Column B(Email) → Column C(Amount)
[Actual Sheet Structure]
Column A(Customer Name) → Column B(Hidden) → Column C(Hidden) → Column D(Email) → Column E(Amount)
Three Things We Learned
**1. Automation design should be based on data structure, not visual appearance**
When you hide columns just because it looks cleaner, AI gets confused. Hidden columns are human convenience, not an automation guarantee.
**2. Tidiness can be a trap**
We hid intermediate calculation columns because "it looks neater." But this undermined our automation agent's reliability.
**3. Reading Claude Code's logs is the fastest debugging method**
Once we saw what data Claude Code was actually reading, the cause became clear.
What Changed After
Now we never hide critical information. Instead, we split our sheets into two.
• Working Sheet: All columns visible. Agents only read from this.
• Display Sheet: A neatly organized version for users. Humans only see this.
Within the same Google account and workbook, this simple separation reduced errors by 90%.
Small Conclusion
When using AI, never assume that "what I see on screen" equals "what AI reads as data." Hidden things create hidden problems.