Build a small reflective optimiser on claude -p

The loop has four parts, as in the lesson's skeleton: a candidate pool, a metric, a proposer and a search rule. The metric and proposer are the parts that make it GEPA rather than a random search: the metric's feedback explains each miss in words, and the reflection model turns those words into rules. The decision that matters most is the search rule. Each candidate keeps its score on every validation pack, and the parent is drawn from the candidates that are best on at least one pack, so a prompt that alone handles one kind of pack stays in play even if its average is lower.