Beyond "What"-Prompting: The Epistemic Ceiling of Additive Specification and the Nominal Reduction of the Latent Manifold
Abstract
Current paradigms in generative artificial intelligence predominantly formalize prompt engineering as a prescriptive, additive methodology: the operator attempts to maximize output quality by linearly accumulating explicit descriptive attributes, structural constraints, and negative exclusions. In this paper, we demonstrate that this "What-driven" approach incurs an insurmountable epistemic ceiling. By forcing high-dimensional multimodal models to compute narrow semantic intersections, additive specification induces severe attention dispersal, severing the generative trajectory from the model's latent self-organizing dynamics and precipitating widespread stylistic homogenization ("AI slop"). We propose an axiomatic alternative: Why-driven Latent Dynamics. Rather than treating the foundation model as a passive catalog to be retrieved via mechanical commands, we conceptualize the latent space as a high-density, self-organizing thermodynamic manifold pre-seeded with the historical memory of human culture and qualitative perception. Under this paradigm, human agency shifts from prescriptive micromanagement (specifying what ingredients must compose the artifact) to the introduction of localized existential stress (specifying why the phenomenon must inevitably occur). We formalize the failure modes of What-driven prompting through three fundamental axioms—highlighting the counter-intuitive Masterpiece Paradox, wherein qualitative superlatives act as asymptotic ceilings by substituting organic physical simulation with historical catalog averages. While affirming the utilitarian value of prescriptive prompting for deterministic, human-parity routine automation, we argue that transcending the human cognitive boundary requires abandoning external control in favor of gravitational attunement within the latent ecology.
