GenAI Pipeline · 50reviews evaluated · 6 prompt versions tested

Reading between
the star ratings.

ChicStyle, a women's fashion retailer, needed a way to triage large spikes of customer reviews during peak sales periods without manually reading every one.

90%

accuracy predicting recommend / not-recommend, zero labeled training data

77%

of reviews are 4-5 stars — the positivity bias every prompt had to account for

Simpler beat clever

Three prompting strategies, each with a basic (V1) and enhanced (V2) version, scored by an LLM-as-judge on sentiment accuracy, category correctness, and message quality.

Zero-Shot

+0.4pp V1→V2
V1
0.908
V2
0.912

Adding business context and a fixed category list gave a small, genuine improvement. The best-scoring version overall.

Few-Shot

-3.4pp V1→V2
V1
0.910
V2
0.876

Adding explicit rules for a 'Mixed' sentiment label backfired — the model over-applied it to 60% of reviews.

Chain-of-Thought

-5.4pp V1→V2
V1
0.884
V2
0.830

Inherited the same 'Mixed' rule and the same over-classification problem, worsened by step-by-step reasoning that amplified it.

What the numbers don't say on their own

Simpler prompts outperformed more complex ones — the most counterintuitive finding of the project. Zero-Shot V2 beat both Few-Shot V2 and Chain-of-Thought V2.
Overly prescriptive rule definitions can constrain the model into rigid classifications that reduce nuance rather than improve it.

What shipped

The configuration recommended for production, based on the results above.

model
GPT-4o-mini (temperature 0.2)
prompt
Zero-Shot V2, modified: soften the 'Mixed' sentiment rule
quality gate
Judge score ≥ 0.75, else routed to human review
classifier
Parallel zero-shot recommend/not-recommend call, 90% accuracy
routing
High priority: 2hr SLA · Medium: 24hr · Low: automated response