Codex 101

A 10-minute orientation to OpenAI Codex — written by a 30-year B2B sales guy who is teaching himself the AI toolkit out loud.

If you have used ChatGPT but you have not touched a coding agent yet, start here. If you already write code for a living, this will be too basic — go read the official docs.

---

What Codex is, in one breath

Codex is OpenAI's coding agent. You give it a goal — "add a login page," "write tests for this file," "fix the bug in line 47" — and it goes off, edits the code, runs commands, and hands you back a pull request. It runs on GPT-5.5.

The important word is agent, not chatbot. ChatGPT talks to you. Codex does the work.

Why I bothered writing this

I am pivoting from traditional B2B sales into AI sales. Every AI sales role on LinkedIn wants someone who has "actually used the tools." So I am using them. This repo is one of those tools, documented in plain English, the way I wish someone had explained it to me three weeks ago.

If you are also a sales person, a PM, a founder, a recruiter, or anyone who has been ChatGPT-curious but has not crossed into the agent world — this is for you.

The four places Codex lives

OpenAI ships Codex on four surfaces. Which one you pick matters more than people admit.

Surface	What it is	Best for
Codex App	Standalone desktop app	First-time users. Visual. Easy to see what it is doing.
IDE extension	Plugin for VS Code, JetBrains, etc.	If you already write code in an editor.
CLI	codex command in your terminal	Power users. Fastest feedback loop.
Codex Mobile	iPhone / Android app (May 2026)	Kicking off long-running tasks from your phone. New.

If you are starting cold, install the app. You can graduate later.

First 5 minutes

# Install (Mac, via Homebrew)
brew install openai/tap/codex

# Log in with your ChatGPT account
codex login

# First prompt
codex "write me a python script that prints hello world"

That is it. Codex will show you the file it plans to create, ask for approval, write it, and (if you let it) run it.

Three rules I learned the hard way:

1. Always start in a new folder. Codex edits real files. Do not point it at your dissertation.
2. Read the diff before you approve. It will ask. Saying "yes" without looking is how things break.
3. One job per session. Long sessions drift. New folder, new goal.

Goal Mode (the thing everyone is talking about)

Until recently Codex was turn-by-turn — you ask, it does, you ask again. Goal Mode went generally available in May 2026 and changes the rhythm: you give Codex an objective, walk away, and it works on it for hours or days, pinging you when it needs a decision.

In practice this means you can tell Codex "ship a working version of this app" before lunch and check in after dinner. It is not magic — it makes mistakes, it gets stuck, it sometimes does the wrong thing well — but it changes what one person can build in a day.

This is the feature that makes Codex feel like a junior dev instead of a fancy autocomplete.

Codex vs Claude Code vs Cursor (the honest version)

You are going to hear all three names. Here is when each one is right.

• Use Codex when you want to delegate a task and walk away. Async work. PR-shaped output. Goal Mode shines here.
• Use Claude Code when you need to think with the agent on a hard problem — refactoring a big codebase, debugging something weird. Opus 4.7 still leads on architectural reasoning.
• Use Cursor when you are at a keyboard, writing code yourself, and want the agent to assist as you type. IDE-native autocomplete + chat.

None of them is "the best." They are different shapes. Most people I know who code for a living end up using two of the three.

What it costs

As of May 2026:

• Codex is included with ChatGPT Plus ($20/mo) for most usage, with metered pricing past that.
• Claude Code Pro is around $17/mo annual.
• Cursor Pro is $20/mo.

If you are evaluating these for a team, the real cost is not subscription — it is how much you actually use them. Run a pilot, count tokens.

Five gotchas I hit in the first week

1. It will happily delete files. Always run in a folder under version control (git init) so you can undo.
2. The cloud sessions are separate from the local CLI. Goal Mode runs in OpenAI's cloud sandbox. Your local files are not there unless you push them up.
3. "Approve all" is a trap. Read every diff for the first month. You learn faster.
4. It bluffs about library versions. If it says "use foo@2.0," verify. It is often a month behind.
5. Cost adds up on long-running goals. Goal Mode can burn through tokens overnight. Set a budget.

What to do next

If you read this far, the highest-leverage next step is:

1. Install Codex.
2. Open the examples/01-first-prompt.md file in this repo and run through it.
3. Come back here and tell me what was confusing — open an issue, I will fix the doc.

Who I am

Parker Robinson. Thirty years in B2B sales, now pivoting into AI sales. I am building in public on LinkedIn at linkedin.com/in/parkerrobinson. This repo is part of that.

If you found this useful, the most useful thing you can do is share it with one person who is "AI curious but not technical." That is who I wrote it for.

---

Last updated: May 25, 2026. Codex moves fast — if anything here is stale, open an issue or PR.