How would you design a system using GenAI to clean up a codebase of 2 million lines across 14 microservices, focusing on what to modify and what to avoid?

1. Decompose the cleanup request into scoped tasks (formatting, dead code removal, API updates). 2. Build a classification layer to mark files as 'never touch' based on criteria (auto-generated files, security-critical modules). 3. Ensure changes are gated by test coverage (zero coverage = flag for review). 4. Map dependencies across services to prevent breaking the system. 5. Create small, reviewable pull requests (PRs). 6. Validate changes before proposing (run tests, linters). 7. Tier confidence for auto-apply versus review based on change type.