# data-pipeline — CLAUDE.md ## 1. Purpose ETL pipeline for customer analytics. AI operates autonomously with safety rails. Human review required only for: schema changes, production deploys, new data sources. ## 2. Constraints - NEVER modify migrations/ without explicit approval - NEVER drop tables or delete production data - NEVER skip tests. If tests break, fix them before proceeding. - All database operations must be wrapped in transactions ## 3. Code Style - Python 3.12, type hints required on all public functions - Comments explain "why", not "what" - f-strings over .format() - Errors: raise specific exceptions, never bare except ## 4. Git Rules - Commit messages: "type: description" (feat/fix/refactor/test/docs) - Pre-commit: ruff + pytest - NEVER: force push, reset --hard, amend published commits ## 5. Quality Gates Before claiming any task is "done": - Code compiles / passes syntax check - Tests pass (or new tests written) - Changes reviewed via git diff - No unintended side effects ## 6. File Structure - AI can modify: src/, tests/, scripts/ - Off-limits: migrations/, config/prod/, .env - New files go in appropriate src/ subdirectory ## 7. Recovery Protocols - Task fails 3 times: stop, document in pending_for_human.md - Unsure about destructive action: ask before proceeding - Test breaks after changes: fix before moving on