Effective AI pair programming requires knowing when to lead and when to follow. Mental model: AI as a very capable intern - excellent at implementation, needs clear specifications and oversight. Tasks where AI leads: boilerplate code, CRUD operations, standard patterns, documentation, test generation for well-defined functions. Tasks where you lead: architecture decisions, novel algorithms, security-sensitive code, business logic with complex rules. Task decomposition for AI: break down to single-responsibility functions. "Write a function that takes a list of product objects and returns the top 5 by revenue, handling null revenue values by treating them as 0." Verification practices: never commit AI code without understanding it. Read through generated code, ask AI to explain any part you do not understand. Run tests, check edge cases manually. Review diff carefully before committing. Maintaining skills: solve problems yourself first before asking AI. Use AI to check your approach rather than generate it. Keep a "no AI" mode for skill maintenance. Code ownership: you are responsible for the code in your codebase. "The AI wrote it" is not a defense for security vulnerabilities or bugs. Review standards: treat AI-generated code with same scrutiny as external code review. Refactoring: AI-generated code often needs style normalization, variable renaming for clarity, extraction of reusable utilities.