The Problem: Manually Sorting Addresses Every Time an Order Arrives
As a non-developer team running a B2B business, the most repetitive task was this.
Buyer order notes kept streaming in. Some wrote in English, some in local languages, others mixed abbreviations. Address formats varied wildly. Someone had to manually read these notes every day and sort them into a Google Sheet. It seemed trivial, but errors were frequent and time-consuming.
"Can Claude Code learn to do this?"
Experiment Step One: Prepare Training Data
First, we copied 30 existing order notes into a Google Sheet (already correctly classified). We fed these to Claude Code as learning samples.
Input: 「Bangkok, Rama 9 Rd, TH」
Output: shipping_city(Bangkok), shipping_address(Rama 9 Rd), shipping_country(TH)
Input: 「서울 강남구 테헤란로 123」
Output: shipping_city(Seoul), shipping_address(Gangnam-gu Teheran-ro 123), shipping_country(KR)
Claude Code recognized the pattern. But initially, it couldn't interpret abbreviations. It didn't know "SG" meant Singapore.
Failure and Correction
When we auto-classified 15 orders on the first try, accuracy was only 73%. We reviewed the error logs.
• City and address confusion: 5 cases
• Failed to recognize country code abbreviations: 4 cases
• Failed parsing when no comma separator: 3 cases
"We need to make it smarter."
We added 20 more samples. We focused especially on abbreviation handling and comma-less formats. At the same time, we instructed Claude Code: "If unsure, mark as 'unclassified'." Safety came before perfection.
Two Weeks Later
Using the same 30 orders for re-testing, accuracy jumped to 91%.
What surprised us more was speed. Manual sorting of 30 orders took about 25 minutes. Claude Code finished the same job in 2 minutes.
We didn't pay the bot overtime, but it didn't refuse to work late (joke intended).
This Part Actually Mattered Most
We didn't trust the automation 100%. We ran weekly sample audits. Anything Claude Code marked as "unclassified," we reviewed by hand. During this process, we discovered patterns.
• Cases where city names changed (e.g., ordering using a city's old name)
• Abbreviations that referred to building names, not city names
• Entries with only postal codes, no city names
We classified the things Claude Code "couldn't" handle, then fed those back into the training data. This loop repeated, and accuracy steadily climbed.
One Month Later
Accuracy reached 96%. The remaining 4% was mostly genuinely messy data. Even humans would hesitate.
More importantly, our team's time spent on address sorting dropped by 80%. Work that once took 2 hours per week now took 15 minutes.
What did we do with that time? More meaningful work. We read buyer feedback, researched new markets, and sat down with the team to plan the next month over coffee.
The Secret: Repetition Over Perfection
The real reason this automation succeeded wasn't Claude Code's brilliance. It was that we "accepted failure and kept feeding back corrections." If we'd given up after the first mistake, this system wouldn't exist.
If you have similar work, try this approach.
1. Gather 20-30 accurate examples
2. Teach Claude Code the pattern
3. Test with small batches (10-20 items)
4. Analyze failures
5. Update your training data
6. Repeat
That's our method.