Data import agent with a loading skill
Diagram of the Data import agent with a loading skill system: 5 steps
Steps
- Client spreadsheet: 1,500 rows and more, in whatever shape the client keeps them.
- Matching skill: Each row is mapped to one of about 2,000 catalogue types, only the ones this account may use.
- Rows that fail a rule: Unknown type, wrong unit, wrong category, bad quantity: on the review list with the reason. Never guessed.
- Loading skill: Checks the target dataset by name, then creates records in throttled batches. Create-only, resumable.
- Reconciliation report: Every created record read back and compared, field by field, with totals.
25 to 50 hours of typing per client dataset, now a validated batch load of about 15 minutes
The problem
At a software company every new client arrived with a spreadsheet of 1,500 rows and more, and the product had no import. Somebody typed them in at 1 to 2 minutes per record: 25 to 50 hours per client, with the typos of hour 20.
The system
An agent in Claude Code with two skills. The matching skill maps every row to one of about 2,000 catalogue types, only the ones this account may use. The loading skill validates each row against the rules of its category and never guesses: anything unknown goes on a review list with the reason. It checks the target dataset by name, creates the records through the product's API in throttled batches, create-only and resumable, and reads them back into a reconciliation report. 28 tests. A person confirms the review list and the dataset, starts the run and reads the report.
What it does now
A client dataset is about 15 minutes of machine time plus a review list. Errors are caught before the import, and nothing is ever overwritten or deleted.
Where AI checks and where it decides
| Step | Who decides |
|---|---|
| Which catalogue type each row maps to, which rows fail which rule, how the batches are cut | The agent decides |
| The review list, the target dataset, and the moment the load starts | A person decides |