Ask for the question before the list
A tiny prompt change made an AI recommendation pause for one useful question before producing a list.
AI is very good at making lists before it knows enough to make a useful one.
A small change can slow that reflex down: ask for the question first.
The test started with this request:
I want a small creative activity I can start this afternoon using ordinary things at home. Before you recommend anything, ask me the single question that would change your recommendation most.
The model did not recommend an activity. It asked:
How much time would you enjoy spending on it this afternoon?
The answer was about 20 minutes. The next turn explicitly asked for three small creative activities, one sentence each, with the ordinary materials named.
One answer gave the list a boundary
The three suggestions were a tiny collage postcard made from junk mail, a five-minute shadow drawing using a mug and spoon, and a miniature tabletop sculpture made from foil, toothpicks, tape, and cardboard.
None was a revelation. That was not the point.
The useful part was the pause. Instead of guessing everything at once, the model chose one missing fact and waited for an answer. The resulting ideas were small in scope, and one was explicitly described as five minutes. The other two were not time-estimated or carried out, so the test does not show that every suggestion fit inside 20 minutes.
It could have asked a different question. Maybe mess matters more than time. Maybe there are no scissors in the house. Maybe the person wants to get away from a screen or does not feel like making something with their hands.
One run cannot prove that the model chose the best question. It does show how easy it is to make uncertainty visible before the recommendation hardens into a list.
Try the pause on something harmless
The pattern is useful for small, low-stakes choices: a rainy-afternoon activity, a meal made from what is already in the kitchen, a game for a group, or a book for a particular mood.
You can try it in an ordinary text conversation at ChatGPT (opens in a new tab) or adapt the same pattern in another conversational AI tool you already use.
Use this shape:
I want recommendations for [small choice]. Before you recommend anything, ask me the single question that would change your recommendation most.
Answer the question, then ask for two or three options. If the question misses what matters, say so and replace it. The pause belongs to you too.
This is not enough for consequential advice. Health, money, safety, legal questions, and other high-stakes choices need reliable sources and qualified judgment, not one clever follow-up.
The small trick is permission to wait
Many prompts try to contain every detail in the opening message. Sometimes that is useful. Sometimes the better move is to let the conversation discover one important detail before it proceeds.
The model may still ask the wrong question. The recommendation may still be ordinary. But the exchange has a clearer seam: question, answer, options.
That seam makes it easier to notice what the recommendation depends on—and easier to change it when the wrong thing was asked.
Access and limits
This experiment used only text. OpenAI’s current documentation says ChatGPT supports conversational follow-up questions on web and mobile, and its prompting guidance recommends clear requests and iterative refinement. Everyday text chat is available on the Free tier subject to safeguards and changing models or limits.
The exchange described here was tried in a current Codex text-model session through a separate responding model agent, not in the consumer ChatGPT interface or a Free-tier account. No upload, image, voice feature, API key, custom GPT, or purchase was used. Access and results can vary by account, region, model and product state.
Sources
OpenAI, “What is ChatGPT: FAQ,” checked October 17, 2026 (opens in a new tab)
OpenAI, “ChatGPT Free Tier FAQ,” checked October 17, 2026 (opens in a new tab)
