ChatGPT and Codex
Chat, delegate, build. The same three moves you learned on Claude. OpenAI has all three: ChatGPT is the chat, Codex is the one you hand a job to.
“If I want a quick question, I honestly go to ChatGPT, because it’s just quick.” Two tools, two jobs. Running both is not indecision.
This is the OpenAI side of the fence. The framework that works on either is AI Fluency →
ChatGPT
The quick question, answered fast
This is the chat. One question, one answer, no ceremony.
It is where I go for a fact, not for a finished piece of work.
- Is this a question, or is it a job?
- Would an answer in the chat window be enough?
- Does anyone else need to see the working?
In the desktop app, OpenAI’s docs say: “Choose ChatGPT or Codex. In ChatGPT, use the toggle above the composer to select Chat or Work.” You can download it for macOS, Windows or Linux.
Codex
Hand over the job, review the result
OpenAI calls Codex “OpenAI’s coding agent for software development tasks”.
You describe the outcome. It reads your files, makes the edits and shows you what changed.
- Can you say what a good result looks like before it starts?
- Is the work in a folder it can actually reach?
- How will you check it, and how will you undo it?
It is a coding agent, and the docs keep it there. The pricing page says “ChatGPT Work and Codex are included in your ChatGPT Free, Go, Plus, Pro, Business, Edu, or Enterprise plan”. Checked 17 September 2026.
The jobs I give it
Repetitive, mechanical, easy to check
In September I needed three questions asked of three AI search engines, three times each.
Twenty-seven runs, one table. Codex did the runs. I read the table.
- Is this the same small task, many times over?
- Could you tell in a minute whether the output is wrong?
- Is anything here live, paid for, or somebody else’s?
The other one: a Search Console setup, written up twice, once for an engineer and once in plain English. What stays in Claude: the thinking, the writing, and anything that has to sound like me.
Computer use
Let it drive the screen
A separate feature in the same desktop app. The mouse moves, the browser opens, the form fills itself.
It is how those search-engine checks got run, with Codex selected. No code, just a machine clicking.
- Could a person do this in a browser in ten minutes?
- Are you already logged in to the thing it needs?
- What is the worst click it could make?
OpenAI: “In supported regions, Computer Use in the ChatGPT desktop app is available on macOS and Windows with ChatGPT Work and Codex.” You install it as a plugin. On macOS it asks for Screen Recording and Accessibility.
A brief, a check, a finished thing
None of this is new. It is the four Ds on a different tool. Codex is stricter about the setup, which is a gift, not a tax.
- 1
Write the brief. Set the stage, define the task, specify the rules. The same three steps.
- 2
Put your standing rules in a file. The docs say Codex “reads AGENTS.md files before doing any work”.
- 3
Take a checkpoint. OpenAI’s own advice: create Git checkpoints before and after a task, so you can revert.
- 4
Read what changed before you keep it. In the terminal,
/reviewreports findings without touching your working tree.
The brief is the whole game, and it travels. Write it here first: How to Prompt →
Four places, one agent
The desktop app
the easy start- Open the ChatGPT dropdown, pick Codex.
- Each chat runs Local, Worktree or Cloud.
The terminal
the CLI- “Terminal client for running Codex interactively or in scripts.”
- You choose the model, effort and permissions.
Your editor
the extension- VS Code, JetBrains IDEs, Cursor and Windsurf.
- Add the open file or a selection to the composer.
The cloud
hand it off- An OpenAI-managed environment, working remotely.
- Connect GitHub, or GitLab in beta.
Local and Worktree chats run on your own computer. Cloud does not, so look at what a repository holds before you connect it.
- ChatGPT: the chat. Quick questions, answered in the window.
- Codex: OpenAI’s coding agent, for work you delegate and then review.
- Computer use: the desktop app driving your screen, on macOS and Windows.
- Where it runs: the desktop app, the terminal, your editor, the cloud.
- The discipline: a brief, standing rules in a file, a checkpoint, a diff you read.
This week: give Codex one job you would not mind redoing, then read what it changed.
The brief is the same on either tool: How to Prompt → What to delegate in the first place: AI Fluency →
Do it with someone. Everything here is free and you can run it alone. If you would rather work through it with me, one to one, book a free 30-minute chat. No pitch.