architecture-drift-signalslisted

# Architecture Drift Signals How AI-generated codebases degrade — and how to measure it. Architecture drift is a **cumulative** failure pattern: no single commit is the culprit, but after 200 prompt sessions the layer boundaries have dissolved. Every individual change looked locally reasonable at review time. This skill gives you the **quantitative signals**, **thresholds**, and **detection commands** to measure drift before it becomes unrecoverable. ## The Mechanism (why AI codebases drift) Three reinforcing root causes: | Code | Root cause | Mechanism | |------|-----------|-----------| | RC01 | **Local optimization** | Each prompt session solves the immediate problem in the most convenient location. The AI has no cross-session memory of which layer a concept belongs in, so business logic lands in UI components, DB queries in route handlers, etc. | | RC02 | **Dependency-graph corruption** | Drift manifests as cycles, cross-layer imports, and hub modules. Once cycles exist, every subsequent prompt has unpredictable blast radius — which accelerates further corruption. | | RC03 | **Structural entropy** | Entropy only increases. The more drift has accumulated, the higher the cost of "doing it the right way", so under pressure the most convenient location wins — again. | **Implication**: drift is self-reinforcing. The earlier you measure, the cheaper remediation is. ## The Three Detection Signals ### FP001 — File-size distribution (the drift fingerprint) Oversized files