Models are bad at “is this code ok.” They are surprisingly good at “is the dependency manifest internally consistent.” The difference is layer-of-concern scoping.
A flat review prompt asks the model to do everything: dependencies, build, network, transport, session, presentation, business logic. The output is shallow because attention is diluted across layers.
Layers of prompts, each scoped tight, run in parallel. Latency drops to the slowest auditor. Findings are layer-attributed and auditable. Cross-layer prompt contamination disappears because the prompts can’t see each other.