Definition
Plain language
A prompt-optimization system that uses language-model feedback to evolve better prompts.
As stated in the literature
A reflection-driven prompt optimizer that mutates and selects prompts based on LLM-judged feedback; specialized baseline subsumed by optimize-anything and outperformed by SkillOpt.
Why it matters: Automated prompt optimization removes a tedious manual loop, and methods like GEPA establish baselines that newer general optimizers have to beat.
For example, GEPA tries dozens of prompt rewrites for a task, has an LLM critique each output, and keeps the prompts that produce the best critiques.
Heard on the show
“Reflexion, Self-Refine, TextGrad, GEPA, a bunch of "EvoSkill" and "AutoSkill" variants.”Episode 078 — Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training