Before I let Codex touch a new codebase, I always do these three things. First, I create an AGENTS.md file. Think of it like an instruction manual for the AI. It contains the project structure, coding standards, frameworks, naming conventions, and anything else I want Codex to remember while it’s working. That one file alone can dramatically improve the quality of the code it generates. Second, I never start by asking Codex to build the app. Instead, I ask it to create a technical plan and wait for approval before writing a single line of code. The planning phase helps uncover problems, missing requirements, and better approaches before time gets wasted building the wrong thing. And third, I make Codex interview me. I tell it to act as a senior product manager and ask questions one at a time until it fully understands the application. The answers often reveal gaps and opportunities that I hadn’t even considered. A lot of AI coding mistakes happen because the AI lacks context. The better the context, the better the plan. The better the plan, the better the code. Context first. Plan second. Code third.