~/ klt.codes
v0.3.2 · Tampa, FL
← all entries
2026-01-21 · 6 min read

Why I made a layered code review

Layers of concern are easier for a model to reason about than a flat 'is this code ok'.

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.

reviewpattern
back to log →