Reflective prompt optimization
A family of methods that improve prompts by scoring, reflecting on failures, and writing better variants.
Reflective prompt optimization scores prompt candidates against examples, studies failures, and writes better variants. GEPA is one approach.
What it means
Instead of editing a prompt by instinct, reflective prompt optimization runs candidate prompts against examples, reads what failed, and uses that feedback to generate stronger prompts. GEPA is one concrete approach in this broader pattern.
Why product teams care
It is often the cheapest rung before training. A better prompt can improve accuracy, reduce retries, and make a workflow more stable without changing model weights.
Understudy angle
Understudy can capture repeated workflows, build evals, and help agents search for prompts that score better before escalating to fine-tuning or routing.
Prompt optimization should be evidence-driven: candidates, examples, scores, failure analysis, and a held-out check.
Take one repeated prompt and score three candidate rewrites on the same examples.