The Accidental Discovery: Can PDFs Really Be Automated?
Last week, while processing buyer documents by connecting Google Sheets and Claude Code on a Windows device, I accidentally included a PDF file path in my script. Naturally, I expected it to fail.
But Claude Code read the PDF.
「The contract's 'payment terms' are within 30 days, and 'shipping terms' are CIF」
It was accurate. It even auto-filled the data into Google Sheets.
The Trap of Automation: Reading Doesn't Mean Understanding
The first 3 to 4 contracts were flawless. The dream team hamsters and puppies celebrated the success. But the 5th file cracked the facade.
Original contract: 「Minimum order quantity: 500 units」
Claude Code extracted: 「Minimum order quantity: 50 units」
It misread the number. One zero was missing. More issues followed.
• When PDFs are scans (images), accuracy drops below 50%
• Signature sections get misclassified as 「key clauses」
• In multilingual mixed contracts, specific languages are selectively ignored
The Non-Developer's Mistake: Overconfidence
The mindset of 「Claude Code did it, so it must be right」is dangerous. Those first 3 to 4 successes were just luck. What's actually needed:
• A verification step for extracted values (manual, not automatic)
• Process only high-confidence contracts for automation
• Set separate alerts for 「critical errors」like amounts and quantities
After testing on both Mac Mini and Windows devices, the environmental differences were minimal. The problem wasn't the tool, it was the process.
So What Now?
I still use Claude Code's PDF processing. But I use it differently now.
Automation zone: Standardized forms, checklist-style confirmation items
Manual zone: Data validation for amounts, quantities, dates
This balance has proven most stable so far. A small epiphany for a non-developer who once dreamed of complete automation.