Master 97% of Codex Features in 3 Minutes: A Practical Guide

Codex has become easier to use, but many treat it as just a prettier chat interface. To unlock its full potential, start with three foundational steps.

Basic Three Steps

1. Create a project workspace – Gather all related files, screenshots, and documentation in a single directory before issuing any large request. A clear project boundary helps Codex stay on target and improves output quality.

2. Write AGENTS.md – This file acts as a project brief that Codex reads at the start of each conversation, reducing the need to repeat background information. Include four key sections: build and test commands, project‑specific coding conventions, hard‑line rules (e.g., "Never submit .env files"), and code‑location strategy.

3. Use plan mode for multi‑step tasks – When a task exceeds two or three steps, enable plan mode so Codex first outlines goals, scope, and risks before executing. This prevents unintended changes, as illustrated by a case where skipping planning caused unrelated files to be modified.

Extended Capabilities

4. Connect your phone – The Codex app can receive tasks and show progress on mobile, turning spontaneous ideas into tracked work without needing a full desktop session.

5. Create Skills – Package frequently used workflows (e.g., writing papers, generating PPTs, weekly reports) into reusable Skills so Codex executes them consistently across platforms like Claude Code and Cursor.

6. Install plugins – Plugins let Codex interact with real tools such as Chrome, Docs, Spreadsheets, Presentations, GitHub, Gmail, Canva, and video tools. Install plugins by use‑case rather than all at once to avoid decision overload.

7. Use MCP (Model‑Connector‑Plugin) integrations – Connect external services (GitHub, Notion, Linear, databases) so Codex can read project data, create issues, and summarize meetings. Start with one reliable integration before adding more.

8. Automate recurring tasks – Schedule Codex to fetch the latest LLM papers from arXiv, filter top scores, and produce a concise summary. Clearly define source, filter criteria, output format, and storage location, and watch for model selection defaults that may affect performance.

9. Learn to read Git diffs – Even non‑programmers should understand what files Codex changes, what is added or removed, and how to revert if needed. Viewing diffs is essential to stay in control of AI‑driven edits.

10. Use Worktree isolation – Create a separate Git worktree for experimental changes, preventing pollution of the main codebase. Choose Local for tiny tweaks, Worktree for substantial features, and Cloud for long‑running automation.

Conclusion

By mastering the three basic steps and progressively adding Skills, plugins, MCP, automation, Git awareness, and Worktree isolation, Codex evolves from a simple chat assistant into a comprehensive AI‑augmented development workstation.