A lot of people open Codex for the first time, type “build me a website,” get back a pile of files they don't understand, and give up. The problem usually isn't the tool, it's how it's being used. Codex isn't a chatbot. It's more like a junior engineer sitting at your computer: it can read your project, edit files and run commands, but you need to explain the task clearly and set boundaries.
Step 1: Pick an interface
There are several ways to use Codex: the command line (Codex CLI), an editor extension, the desktop app, and handing tasks off to run in the cloud from the web. If you're a complete beginner, start with the desktop app or the editor extension. The interface is intuitive and you can see exactly which files it changed. Once you're comfortable, move to the CLI, which is faster and better suited to batch work. Whichever you choose, you can sign in with your ChatGPT account, and it's the same capabilities under the hood.
Step 2: Open it inside a real project
Don't ask it to “build a big system from scratch” in an empty folder. A better way to start is to pick a small real project you already have, even if it's just a static web page or a script that processes Excel files, and launch Codex in that directory. It reads the folder structure first, then acts on your instructions. For your first prompt, try asking “What does this project do, and how do I run it?” and check whether it understood correctly.
Step 3: Keep your first task small and verifiable
A good first task looks like this: “Change the homepage title to XXX, then start the local server and tell me which address to open to see the result.” It has three qualities: small scope, results you can see right away, and a built-in verification step. Once you've confirmed it got it right, give it the next task. Doing one thing at a time works far more often than handing over ten things at once.
Step 4: Use AGENTS.md so it remembers your project
Repeating “we use pnpm,” “the test command is npm test” and “don't touch the config folder” every time gets old fast. Codex reads an AGENTS.md file in your project's root directory, so put your project conventions there and it will follow them every time it starts. A good-enough AGENTS.md only needs to cover how to install dependencies, how to run the project, how to test it, which folders are off-limits, and any code style rules. You can have Codex generate a first draft and then edit it yourself.
Step 5: Don't hand over full permissions right away
Codex can ask for your approval before it edits files or runs commands. When you're starting out, keep approvals on and watch what it wants to do at each step. You'll learn a lot, and you'll catch mistakes before they happen. Once you trust how it behaves in a given project, loosen things gradually. Also, make a Git commit before starting a task so you can roll back any time something breaks. It's the cheapest insurance there is for AI-assisted coding.
The four most common beginner mistakes
First, tasks that are too big: asking it to build an entire product in one sentence, and watching it drift off course halfway. Second, using its changes without reviewing them, so problems pile up. Third, saying “it doesn't work” when you hit an error instead of pasting the actual error message. Fourth, skipping AGENTS.md, so every conversation starts from zero. Avoid these four and Codex becomes much more useful.
If you want to learn it systematically, step by step, from computer basics and setup to the CLI, MCP, Skills and team workflows, check out Zero to Codex (lessons are in Chinese). The first 8 lessons are free and you can watch them without logging in. For more on writing AGENTS.md, read AGENTS.md: putting your project conventions in a file.