Stories have been shown to be potentially valuable for language model pretraining because their semantics are largely self-grounded. Existing human storytelling datasets are relatively small, and existing synthetic datasets lack realism. Here we describe the guided synthesis of a story corpus that is both large and realistic. We evaluate a set of metrics for measuring the quality of individual stories, including coherence, surprise, complexity, as well for quality of the dataset as a whole. Using the metrics we extensively evaluate multiple prompting approaches and three small models (3-8B). Based on the results, we steer dataset generation to yield a high-quality ’synthetic twin’ of a Dutch children’s storytelling collection, similar in terms of lexical and semantic diversity, but much larger, which we call *Stories for the Win* (StoriesFTW). We validate our metrics and evaluate samples from our corpus with crowd, expert, and LLM judgements.