I turned Andrej Karpathy's tips on understanding LLM output into an Open Source Agent Skill