Small-Model Classification Workflow Adds an Explicit Review State
A poster shared a small-model classification workflow that defines labels, requires a structured output, and allows an explicit “review” result for unclear, mixed, or unsupported cases. The poster said they had been testing the pattern with Claude Haiku 5.5 and suggested evaluating classification accuracy, cases sent to review, retry rate, latency, total cost, and human correction time. They also disclosed that the linked Claude Haiku 5.5 resource was their own site. The report contains no independent benchmark results or corroboration of the poster’s testing.
I’ve been testing a simple idea with smaller models:
don’t ask them to solve everything. Give them narrow jobs where the result is easy to verify.
One of the best examples I’ve found is classification.
A bad prompt would be:
“Categorize these support messages.”
That sounds simple, but it creates a problem.