A team is building a structured data extraction pipeline to pull financial line items from 500 PDF invoices per night. After two weeks, the accuracy rate has plateaued at 78% — well below the 95% target. A review session reveals that the team spent most of their development time writing detailed Skill documentation that addresses many scenarios they haven't yet observed in production data. Meanwhile, they have no systematic way to measure which specific extraction failures are most common. What change in development approach would most directly address this plateau?