Introduction to Codex CLI and AI-Assisted Development

Codex CLI reads AGENTS.md from the project root at the start of every session, making our preferences and instructions persistent across conversations. The file can hold coding style, project conventions, and commands Codex should know about, such as how to run tests. Keeping it in the project root means every session starts already calibrated, with no need to re-explain how we work.

In a prompt, we use @ to point Codex at a specific file — for example, @spec.md — so it works from exact context instead of searching the whole project. The ! prefix runs a shell command inline, like !node --test, executing it directly without routing the request through Codex. Together they let us control both the context Codex sees and the commands we run, right from the prompt.

A “next” workflow captured in AGENTS.md tells Codex to find the first unchecked item in our spec file, implement just that item, and mark it complete. Because the steps live in AGENTS.md, we can move through an entire feature list with a single instruction instead of re-explaining the process each session. The spec says what to build; the workflow says how to step through it.

With the interview technique, we ask Codex to question us before it generates any code or spec. Those answers let it tailor the output to our experience level, preferences, and the project’s actual requirements, rather than guessing. It’s especially useful for turning a vague idea into concrete requirements, since the questions surface details we might not have thought to specify.