ChatGPT Keeps Doing Deep Research: Change the Mode

If ChatGPT keeps doing deep research, first check whether Deep research is selected before you send another question. Start a fresh standard chat for a simple answer, and remove instructions that ask for an extensive investigation when that is not your goal.

A long answer alone does not prove a research task started. Look for the research plan and progress interface. Once you find instructions that produce the right workflow, fenbs can keep that reusable prompt outside one conversation so your next session begins with the same scope.

How do you get back to an ordinary answer?

1. Check the tool selected in the prompt area before submitting. If Deep research is active, leave that selection and start a standard chat using the controls available in your app.
2. Remove any research command from the prompt, including /Deepresearch. Paste only the question you need answered, with a clear limit such as one paragraph or three bullets.
3. Try a low-stakes question in a fresh conversation. For example, ask for a plain-language definition using only general knowledge, without a research report.
4. Compare the result with the original conversation. If the fresh chat works, review the old prompt and its instructions before reusing it. If it still starts research, record the behavior for troubleshooting.

OpenAI lists the prompt command, tools menu, and sidebar as ways to start deep research↗ (opens in a new tab). Checking those entry points is more useful than repeatedly telling a running research task to be brief. A request for fewer words concerns the output; selecting the intended workflow concerns how the work starts.

Is research actually running, or is the answer just long?

Separate the visible behavior from your interpretation. Note whether you see a proposed plan, a research progress view, or only a lengthy response in the chat. Write down the exact labels rather than treating every pause, search result, or explanation as the same mode.

That distinction changes the fix. An oversized ordinary answer calls for a tighter output request. A selected research tool calls for a tool change. A page that never loads calls for technical troubleshooting. You can waste several attempts if you treat all three as a single problem with your prompt.

If you are choosing between an extended investigation and a reasoning-focused answer, the separate deep research versus thinking guide covers that choice. Here, the aim is to make the mode match the question you already intended to ask.

What should you change in the request itself?

Describe the smallest useful result. Instead of asking for everything about a market, ask for three questions to discuss with a colleague. Instead of asking for an exhaustive review of your draft, ask for the two sentences that are hardest to understand. These are different assignments, even when the subject stays the same.

Try: “Give me a short explanation from general knowledge. Do not start a deep research task. If current evidence is necessary, tell me what needs checking before investigating.” This is an instruction about your preference, not a guarantee that a prompt can override the selected tool or the app’s controls.

Keep a current-facts exception. If your question concerns a changed policy or a product feature, a quick unsupported answer may be worse than taking time to verify it. Ask for a narrow check or do the check yourself. The goal is appropriate effort, rather than eliminating sources from every response.

For assignments that really require a report, use the separate deep research prompt guide. Save the quick-answer version in fenbs Skills and prompts, with a note explaining when to use it.

Could saved instructions be making the request broader?

Read the instructions you supplied earlier. A sentence such as “always investigate every claim thoroughly” may conflict with today’s request for a quick explanation. Treat that as something to test, rather than proof that it automatically selects a particular tool.

OpenAI lets you edit or disable ChatGPT custom instructions↗ (opens in a new tab). Review any standing preference that asks for comprehensive research or lengthy reports. Change one instruction at a time so you can identify which change helps, and keep a copy of any wording you want to restore.

In a work setting, separate a preference from an approval rule. “Keep everyday answers short” is a writing preference. “Ask me before beginning an extensive investigation” is a workflow instruction. Neither should imply that unverified facts are acceptable. Store the reusable wording and its intended use in the same place.

How can you tell whether the problem is fixed?

Use one repeatable question and change only one thing per attempt. First test the selected tool, then the prompt wording, then your saved instructions. Keep the account and question consistent while comparing results. An unrelated new question makes it harder to learn what actually changed the behavior.

Record the outcome as an observation: “Fresh standard chat answered directly; older conversation opened a research plan.” Do not turn that into a claim that every old conversation is broken. If a simple test unexpectedly starts research across fresh chats, capture the selected controls and exact steps before contacting support.

For a broader process once the mode is correct, see the ChatGPT deep research workflow. The AI at work hub covers the surrounding habits of keeping useful instructions and project context.

How to do this with fenbs

1. Create a free account and add the project whose everyday questions keep turning into research assignments.
2. Add two entries in Skills and prompts: a short-answer request and a research request. Include an example question and the intended output for each.
3. Record any instruction a person has agreed should always apply in Decisions and rules. For example, ask before expanding a quick question into a large investigation.
4. Connect a supported assistant using the setup guide on Integrations. Ask it to read the project context and the relevant prompt before working, then save which version you tested and what happened.

fenbs is free for one person, no credit card. It does not change ChatGPT’s selected mode or control a running research task. If you only need to unselect a tool once, the ChatGPT controls are the right place to finish; fenbs helps when the instructions need to survive repeated sessions or a change of assistant.