Speaking at Every's Think Week workshop, Mike Taylor, head of evals at Every, presented a method for teaching AI to write in a person's voice by building a test rather than adjusting prompts. The method splits 105 of his published pieces into training data, then has two Claude sessions race to match a real article while a judge picks the closer one. The winning skill, optimized against that score, beat a baseline roughly 80% of the time on a 54-piece holdout exam and cost about $50 in tokens to hill-climb over roughly an hour. Taylor said 10 good pieces is the floor and 50 is where it gets "juicy," and that slop in equals slop out. The report also covers which models matched his voice (Sonnet edged Opus, but he still writes with Opus), how the loop draws on Andrej Karpathy's "auto research" idea and DSPy's GEPA optimizer, why human editors survive as the audience's advocate, and the unresolved tension that a skill trained on your current writing locks in your present self. The eval strips a real published piece to a style-free brief, and two Claude sessions write from that same brief, one using the current skill and one a challenger skill, with a judge asked only which output is closer to what Taylor actually published.