I freaking love Claude Code skills—but let’s be real: they’re noisy. For me, about 70% of the time a skill gives the intended output, and ~30% of the time it’s a bag of rocks. In this video I show you how I’m combining Claude Code skills with Andrej Karpathy’s new “autoresearch” idea to make skills dramatically more reliable and accurate—and to literally have them improve themselves overnight. I walk you through the core autoresearch structure (small repo, a “thing to change,” and a loop that measures improvements) and then I translate that to skills: the “thing to change” is your SKILL.md prompt, and the key is building an eval suite with objective, binary tests. I demo this with my diagram-generator skill, where I define four yes/no criteria (legible text, pastel palette, linear layout, and no numbers/ordinals). Then I have an agent generate 10 diagrams every 2 minutes, score them out of 40, mutate the prompt, and keep the winner. The big takeaway: if you can define the right evals, you can auto-optimize basically anything—skills, websites, cold email, landing pages, thumbnails—without vibes, just numbers.