ChatGPT Work Isn't For You

Power users are annoyed. They're also missing the point.

ChatGPT Work is not for power users

ChatGPT Work is probably not for you.

If you're reading an AI newsletter, know what Codex is and have spent any time poking around Claude Code…you are already in the loud, weird power-user minority.

And that minority is currently furious with ChatGPT Work.

Work looks like a watered-down Codex. The interface is messy. The names are rubbish. Some options appear on desktop, some on web, some on mobile, and we have all decided this is a personal attack.

All fair!

But…OpenAI is aiming Work at the other one billion ChatGPT users. People who use ChatGPT every day but will never download a terminal, open Codex or learn what an MCP is.

It’s not for YOU as a power user. It’s for the rest of humanity.

The loud minority

Last week I wrote about why ChatGPT Work matters. It gives normal ChatGPT users access to the agentic stuff power users have been hammering for the last six or seven months. That (for me) is pretty cool.

Those of us who are power users got used to AI doing the work very quickly. That’s just what AI does right?

But most people still ask ChatGPT a question, get an answer and then go away to do the actual job themselves. ChatGPT Work is the bridge from asking to delegating.

So yes, Codex is more powerful. It has deeper controls. It can build and run “proper” software. It’s great. I use it daily for faaar too long. If you're already happy there, stay there!

Just know that Work is the easier door for everyone else.

One app, three doors

Three doors into ChatGPT: Chat, Work and Codex

Admittedly it’s a little confusing now. We have three different ways to use ChatGPT.

The easiest way to choose is to decide what you want back from it.

Use Chat when you want to think something through. Ask questions. Compare options. Research an idea. Argue with it. Get clearer on what you actually want. The chat itself is the output.

Use Work when you already know the outcome and want a finished thing back. A report. A presentation. An analysis. A research pack. Something you can open, check and use. The output will be a file or some sort of artefact.

Use Codex when the finished thing is software. A site, app, database, automation or technical system.

The ten-second version:

  • Think: Chat

  • Make: Work

  • Code: Codex

There is no "best" door. Depends what you need back.

Chat first, then hand it over

And we can chain them!

For example I used Chat to work out the argument for my livestream presentation. We went back and forward until the structure was right. Then I copied the conversation link into Work and told it to build the deck.

Literally shared the chat. Copied the link. Moved to Work mode. Pasted the link and said “let’s work on this”.

Chat helped me decide then Work made the “thing”.

That handoff is useful because giving an agent a fuzzy brief just lets it produce polished rubbish without bothering you for a while. More autonomy does not rescue confused thinking. It merely lets the confusion run for longer without you knowing!! Whilst burning up your usage.

So…start in Chat when you are unsure. Work out the audience, the outcome and what good looks like. Then send the settled thinking to Work to get the “thing” done.

Give Work one proper job

Give ChatGPT Work one proper job

Work can do, well, work. BUT don’t overburden it.

If the task is TOO big then Codex is going to be a better tool.

Instead with Work stick to relatively discrete, simple jobs to start.

For example: pick one job you already understand. Maybe turn a meeting transcript into actions and owners. Compare three supplier quotes. Turn an approved outline and source material into a presentation. Build the weekly report in the same format as last week. These sort of jobs with a defined input and output. Something you can look at and think “yup, it did a good job!”

Give it:

  1. The outcome you want.

  2. The source material.

  3. The boundaries.

  4. What "done" looks like.

Your prompting skills still matter because prompting is communication. The Prompting Fundamentals playbook applies even more when the AI can run off for 12 minutes and do a load of work without you. You still NEED to communicate what you actually want if you want good results.

Oh and as always check the result. Sources, numbers, links, the lot. A nice-looking deck can still be full of nonsense. Check the work!

So. If you are a power user your task today is to keep moaning about ChatGPT Work and the new app! 😛 

Everyone else should open it and give it one real job.

To the Task,

Kyle