We introduce a programming paradigm optimized for the cognitive architecture of Large Language Models, rather than human authors. Traditional languages rely on sequential reasoning and implicit context—abstractions that conflict with the pattern-completion nature of LLMs. SPELL is an AI-native dataflow language with explicit dependencies, explicit types, and structured JSON format. By aligning syntax with the generation mechanism of the model, we demonstrate that the optimal abstraction for AI is fundamentally different from the one for humans.