ChatGPT Chat or Work? Choose the Right Mode for One Office Task

Evidence label: documentation-based guide with an illustrative example. The sample request and outputs below are fictional. No ChatGPT execution is claimed.

ChatGPT now gives you a simple choice that can still cause a messy result: Chat or Work?

The useful distinction is the task, not the label. Chat is designed for quick conversational help. Work is designed for longer, multi-step work and finished deliverables. If you choose the wrong surface, you may spend more time steering the task than doing the work.

This guide gives you a small decision card you can reuse before starting a research brief, report, or other office deliverable.

The quick decision

Choose Chat when you need one of these:

• An explanation of a concept.
• A quick rewrite or brainstorm.
• A short answer based on a small amount of context.
• A conversation where the output may change as you think aloud.

Choose Work when you need all or most of these:

• A defined deliverable, such as a report or comparison table.
• Several steps, files, or source links.
• Clear rules for missing information.
• A reviewable output that someone else will check before it is shared.

Use Codex, not Work, for software development, repository changes, tests, or commands. OpenAI documents Codex as the experience for technical and coding work.

Availability depends on your account, plan, workspace settings, and rollout access. If Work does not appear in your account, use Chat with a smaller scope or follow your organization’s approved alternative.

A fictional task

You manage operations at a 40-person company. Your manager sends this message:

“Can you look at the three supplier proposals and tell me which one we should pick? Please make it presentable for tomorrow.”

The input is incomplete. You have three fictional files:

• northstar-proposal.txt: annual fee $18,000, implementation in 6 weeks, phone support included.
• maple-proposal.txt: annual fee $15,500, implementation in 10 weeks, phone support listed as an optional add-on.
• orbit-proposal.txt: annual fee “to be confirmed,” implementation in 4 weeks, support terms not stated.

The right deliverable is not a confident winner. It is a source-linked comparison with missing terms and questions for the supplier.

This task fits Work because it has multiple inputs, a defined output, and review points. A quick question such as “What should I include in a supplier comparison?” fits Chat.

Three steps to make the choice

1. Write the deliverable before opening a chat

Use this short decision card:

Task: Compare three supplier proposals.

Use Chat if: I need a quick explanation, rewrite, or first-pass idea.
Use Work if: I need a source-linked table, several files reviewed, and a handoff-ready brief.
Use Codex if: The task requires code, a repository, tests, or terminal commands.

Required output:
1. A comparison table with supplier, annual fee, implementation time, and support terms.
2. A source file for every extracted value.
3. A missing-information list.
4. Questions that a human should answer before choosing a supplier.
5. No automatic recommendation when a required field is missing.

Review owner: Operations manager.
Approval rule: Do not send a supplier reply or make a purchase decision.

If you cannot name the output and the review owner, start in Chat and clarify the task first. Work is not a cure for a vague brief.

2. Start the smallest suitable surface

Open ChatGPT and choose Chat or Work according to the card. The exact controls and availability may vary by account and workspace. Do not assume a menu item exists because another account has it.

Attach only fictional or approved files. Then paste this prompt:

Prepare a source-linked supplier comparison from the attached files.

Deliverables:
1. A table with supplier, annual fee, implementation time, and support terms.
2. A source file beside every value.
3. A separate list of missing or ambiguous terms.
4. Five questions for the operations manager to answer before choosing.
5. A short handoff note that says what is known, what is unknown, and what must be checked.

Rules:
- Use only the attached files.
- Copy values exactly. Do not calculate a total unless the files provide the inputs.
- Treat "to be confirmed" as unknown.
- Do not guess missing support terms.
- Do not rank or select a supplier.
- Keep the output under two pages.
- Put the source filename in brackets after each extracted value.

Before drafting, show a short plan and list any file you cannot read.
Stop after the plan so I can approve the scope.

The prompt asks for a plan first. In Chat, a deep research task can also produce a proposed plan for review. In Work, the documented flow asks clarifying questions after the initial prompt. Either way, review the scope before accepting a long run.

3. Check the artifact before you share it

For the fictional inputs above, an acceptable illustrative result would include:

Supplier	Annual fee	Implementation	Support	Source
Northstar	$18,000	6 weeks	Phone support included	northstar-proposal.txt
Maple	$15,500	10 weeks	Optional add-on	maple-proposal.txt
Orbit	Unknown	4 weeks	Not stated	orbit-proposal.txt

The result should also say that Orbit cannot be compared on price or support yet. That is an illustrative answer written by the editor, not a model result.

Run these acceptance checks:

1. Every table value matches one of the three fictional files.
2. Every unknown field stays visible instead of being filled with a guess.
3. The source filename appears beside each extracted value.
4. No supplier is ranked while required terms are missing.
5. The handoff note separates facts, unknowns, and questions.
6. The document does not claim that a supplier was contacted.

A polished table that fails any of those checks is not ready to share.

Common failure fixes

You only see Chat, not Work. Check your plan and workspace access. Work availability is account dependent. If access is not available, use Chat for a smaller, clearly bounded draft and keep the same acceptance checks.

The output is a recommendation instead of a comparison. Repeat the rule that missing terms block ranking. Ask for the source-linked table and the unresolved questions first.

The model invents a support term. Mark the field as unknown, point to the source file, and ask for a corrected row. Do not accept a smooth sentence as evidence.

The task keeps expanding. Use a fixed deliverable, a short page limit, and a stop after the plan. If the scope still grows, start a new chat with only the narrowed request.

TRY, SKIP, USE

TRY Work for a multi-file report, research brief, or comparison where you can define the output and review it.

SKIP Work when the task is a two-minute explanation or a quick rewrite. Chat is probably simpler.

USE Chat for the first clarification when the request is vague, then move to Work only after you have a deliverable and acceptance checks.

For more on the current distinction, read ChatGPT Work and Codex and Deep research in ChatGPT. For more practical AI workflows, get the Weekly Verdict.