You point an AI agent at a vague ticket, it confidently rewrites three files, and now half your tests are red and you are not sure what it even changed. The fix is rarely a better one-shot prompt; it is a repeatable workflow that keeps you in the driver’s seat.
A few core patterns have proven consistently effective across building features, refactoring legacy code, and squashing bugs. They apply equally whether you drive Cursor, Claude Code, or Codex, the only difference is which plan-mode and review surface each tool exposes.
The most successful and reliable pattern for working with an AI assistant on any non-trivial task can be broken down into three phases:
1. Plan Phase: The Architect
In this phase, you act as the architect and the AI acts as your research assistant. Your goal is to explore the problem, understand the codebase, and create a detailed, step-by-step implementation plan. This is a low-risk, high-leverage activity that sets the stage for success.
2. Execute Phase: The Supervisor
Here, your role shifts to that of a supervisor. You delegate the implementation of the plan, one task at a time, to the AI. You review the code it produces, provide feedback, and ensure the implementation stays on track.
3. Verify Phase: The Quality Engineer
Finally, you become the quality engineer. You work with the AI to write tests, debug issues, and verify that the implementation meets all requirements. This ensures that the speed gained from AI assistance doesn’t come at the cost of quality.
This “Plan, Execute, Verify” cycle is the foundation for many of the specific development patterns discussed in the following guides, including the core “PRD → Plan → Todo” methodology.
These guides dive deeper into specific implementations of the Plan, Execute, Verify pattern, showing you how to adapt it to your methodology. Start with the PRD to Plan to Todo methodology, the backbone of the rest.
