This resonates deeply. I've seen this exact pattern while building automation systems for clients — the initial AI-generated code ships fast, but when you revisit it 3 months later for a feature extension, you spend more time deciphering than you saved. What helped me was enforcing an "architecture-first" rule: before any AI generates a single line, I write a clear system design doc with module boundaries and naming conventions. The AI then works within those constraints rather than inventing its own structure. Curious — do you think architecture decision records (ADRs) could be a practical guardrail here?