PROMPTS
ChatGPT's Three Modes And 15 Tasks To Try ⚡
Chat, Work, and Codex are three different tools wearing one logo. Five real tasks for each, every one a copy-paste prompt you can steal and fill in, plus the.
You did not imagine it. ChatGPT looks different, and the story going around is backwards.
It did not break into three pieces. On July 9, 2026, OpenAI merged its Codex app back into ChatGPT (their launch announcement has the full story). Two apps became one, so ChatGPT now has three modes under one roof, not three apps. It feels like a split only because the three really do behave like different products.
Here is the one line that sorts all of it. Chat gives you thinking. Work gives you a finished file or a real action inside an app you connected. Codex gives you something that runs. Name which one you want before you type, and you are ahead of most people.
Below are fifteen, five per mode, each a copy-paste prompt to fill in with your own details. They work whether you have a 9-to-5, run your own thing, or just want your life more handled. Steal them.
What Each Mode Is Actually For
The map first. The last column is the whole decision: what kind of thing do you want handed back?
| Mode | Where you find it | What comes out |
|---|---|---|
| Chat | Left side of the toggle inside ChatGPT | Thinking. Answers, arguments, drafts, a second opinion. |
| Work | Right side of the same toggle | A finished file, or a real action in a connected app. |
| Codex | The other option in the product selector, one level up | Something that runs. A site, a script, a change to your files. |
Work and Codex are two faces of one engine. You are not picking a smarter tool, you are picking the shape of the result: a document, or something that runs.
5 Tasks For Chat Mode
Chat is for what happens in your head. Nothing gets saved or sent, and that is the point. This is where you think out loud with something that actually pushes back.
Make it argue every side and name the risk you are underrating, instead of rubber-stamping what you already want. A job move, a big purchase, a business call.
The hard email or awkward text you have rewritten five times and still might send wrong. Get it clear and kind in one pass.
A raise, a client negotiation, a hard talk with someone you love. Have the bad version first, with nobody watching.
Dump everything you are carrying and get back a sorted, prioritized week. A job, a business, or just life.
Turn any intimidating topic, a contract clause, a health term, a new tool, into the one idea it hangs on.
5 Tasks For Work Mode
Work is the one to actually learn. It is on every plan and it hands you a finished thing, not a wall of text. The test for whether a task belongs here: would you have opened another app to finish it?
Not notes you turn into a doc at 11pm. The actual brief, deck, or report, with sources and a recommendation.
The export nobody wants to touch. Get back an editable spreadsheet you can trust, with every change explained.
Set it once. Every week it pulls what matters and hands you a one-page summary. Needs one app connected.
Six files and a notes doc into one clean proposal, in your format, ready to send. The 11pm assembly job, gone.
Sorting is the easy half. The value is the urgent replies, already drafted in your voice, waiting for you to hit send.
Before drafts and actions work: writing to your apps is off by default, and even once on, it defaults to asking you every time. On Business and Enterprise, an admin enables it per app in Workspace settings. If your drafts never appear, that is why.
5 Tasks For Codex
Codex sounds like it is for engineers. It is not, but it has two gates to clear first, so read these before you try anything. You can download Codex and its desktop app here.
One honest note: OpenAI itself points non-coders toward Work, not Codex. Open Codex anyway for one reason: it gets you far enough to prototype something and learn whether it is worth paying a pro to finish.
The 400 files you were never going to rename by hand. Renamed, sorted, done, no code required.
The same tedious hour every month, built into something you re-run in one click. The lowest-risk way in.
A tracker, checklist, or calculator, published to a link your team, a client, or your family can open. Sites needs Plus.
A real page, live, for your service, an event, or a portfolio, with a form built in. Sites needs Plus.
Codex doubles as a tutor. Have it explain your own setup in plain English, then fix one thing.
Four Things Nobody Warns You About
- Work and Codex share one usage pool. They are the same engine, so they draw from the same allowance. A long Codex session on Tuesday can quietly eat the budget you were saving for Friday's report, and nothing warns you at the time.
- Free and Go are real, just small. Work is on every plan, but Free and Go get a small shared allowance, and the desktop app is the most reliable place to find the toggle. If you cannot see it on the web, check the desktop app before deciding you do not have it.
- The agent tasks are paid-only. Agentic browsing and heavier automations work on every paid plan, not on Free or Go. If you are on Free, that is not where to start.
- Computer Use on Windows takes over your screen. It runs on your active desktop, so it cannot work quietly in the background. You watch the cursor move. On a Mac, grant Screen Recording and Accessibility permissions first.
The Claude Version, In One Table
If you use Claude too, the shapes line up almost exactly. Same three jobs, different logos.
| The job | OpenAI | Anthropic |
|---|---|---|
| Thinking with you | Chat | Claude |
| Office work, finished files | Work | Cowork |
| Building things that run | Codex | Claude Code |
The twist worth carrying: Work is on every ChatGPT plan including Free, while Claude's Cowork is paid-only. OpenAI is handing agentic work to everybody; Anthropic is charging for it. I go deeper on the two in ChatGPT Work vs Claude Cowork.
The mistake is not picking the wrong mode. It is doing all three jobs in Chat because that is the one you know, then deciding the new stuff is overhyped. Ask for the file. Ask for the action. Ask for the thing that runs. Different question, different tool, completely different afternoon.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.