Audiobook listeners returning after long breaks need help recovering a story. We created and deployed personalized Audiobook Recaps to address this memory barrier. Production feedback showed that good recaps are not conventional summaries: memory cues, appropriate tone, anticipation, and spoiler avoidance emerged as task-specific quality dimensions.

We introduce a pairwise quality-bootstrapping method that discovers these dimensions while optimizing prompts. Annotators choose between recap pairs and explain why; an LLM induces a structured pairwise judge, replacing the missing pointwise metric in a reflective GEPA loop. From a minimal prompt, the method beat months of manual prompt engineering by 22% in pairwise win rate using 30 annotations and no reference model. Independent evaluation found no summarization regressions and raised spoiler avoidance from 23% to 80%. This offers a general pattern for generative tasks whose quality criteria are initially unknown.