Let AI Teach Me · Opening

Let AI help me find what to ask next

Getting an answer, spotting what's shaky, and writing the next question — those three layers are the split. When you're done, you'll see which layer you stop on, and which sentence to write next.

12 lunches, budget ¥800, delivered to the meeting room Friday at noon. Admin forwarded this to you Tuesday morning, along with the list of colleagues coming to the meeting.

You send headcount, budget, and delivery time to AI and ask it to build a menu. It comes back fast with a combo:

Four meat dishes, three vegetable dishes, plus a staple, about ¥65 a person; half mild, half spicy; and ¥20 set aside for napkins and disposable utensils.

The answer is actually useful. The counts add up, the budget has slack, you swap two sold-out dishes on the restaurant page, and you place the order before 11. Screenshot into the group chat — job done.

Friday 11:58. The boxes just hit the table. A colleague sees a sauce label and asks if there's peanut in it. That's when you learn she is allergic to peanuts. The restaurant confirms one sauce has crushed peanut, and the same wok has cooked peanut dishes. You add an emergency order. It arrives at 12:36. She waited in the meeting room for 38 minutes.

You clocked out because every task on the screen was checked: headcount met, budget not blown, time written down. Allergy never entered the question, so the answer treated it like a normal group lunch. A complete-looking menu made the unasked condition even harder to see.

A usable answer — why can't you clock out yet?

AI finished the problem you handed it. You gave headcount, budget, delivery time, so it optimized around those three. You didn't say who can't eat what, so it has no way to know there's an allergic person in the room.

A usable answer only means it finished the part you already asked. The conditions that actually change the decision — some are in the question, some are still hiding in what you haven't thought to ask.

That's when you can ask AI to switch roles. Pause the recommendations. Have it check which still-unconfirmed conditions this plan depends on, then rewrite the most important one as the next question. In this lunch scene, it should first ask about food allergies, religious diets, or other restrictions, then remind you to verify ingredients and kitchen handling with the restaurant.

The next question turns a vague worry into something you can go ask a person, check a label, or confirm with the restaurant right now. That menu is already orderable. Will you stop at getting the answer, spotting what's shaky, or writing the next question?

Try it once · See which layer you stop on
The same lunch plan — where each of the three layers ends up
Pick where you would stop right now
Pick a layer, see what happens when you order, and what the next sentence should say.
1 · First get a usable answerThe menu already meets what you wrote down
2 · Then mark what looks shakyYou see it still depends on an unknown condition
3 · Turn the shaky bit into the next questionYou know who to confirm what with next
This lunch chat between you and AI
Pick a layer first, then hit Play to see where this order goes.
This is where it shows which layer you stopped on
This will say what this layer got done, and how the next sentence can ask.
Pick the layer you usually stop on, then hit Play to see the whole run.
What extra step each layer takesLayer 1 takes the answer. Layer 2 circles the unknown condition. Layer 3 turns that unknown into a question you can check.
What happens after you write the next questionYou leave the chat window and go back to the room: confirm restrictions with colleagues, then ingredients and kitchen handling with the restaurant.
Teaching sketch: the answers are simplified demo text, not real output from a model.
Stick these three sentences after any answer

Once you have an answer you're ready to use, paste it back with your scene. Send the three sentences below in order, and AI switches from more advice to helping you find conditions you haven't confirmed.

Copy this whole block
  1. First find the conditions that were never written downThis is the plan I'm about to use: [paste the plan]. My scene is: [who, when, what you need to do]. List the conditions this plan depends on that I have not confirmed, ordered by how much they change the outcome. Do not fill in assumptions for me.
  2. Then ask the one that matters mostStart from the most critical condition. Ask me only one question at a time. After I answer, decide whether to follow up or move to the next condition.
  3. Finally turn it into a check you can run on siteFrom my answers, write a list of what still needs verifying. For each item, say who to ask, what to check, and what result lets us keep using this plan.

In the first sentence, “Do not fill in assumptions for me” is the key. Without that half-sentence, AI may quietly add defaults like “nobody has restrictions,” and the unknowns get covered by a smooth paragraph again.

How a shaky bit becomes a question you can act on

“This menu might have a problem” is only a feeling. At order time you still don't know where to stop. To push it forward, you have to answer three things in a row:

  1. Circle the sentence that would change the decision, for example that this menu assumes everyone can eat everything.
  2. Write down which on-site condition is still missing — here, each of the 12 people's allergies and restrictions.
  3. Name the check and the pass line: confirm person by person, then have the restaurant state ingredients and kitchen handling, and only order once you know those foods can be avoided.

These three layers are an experience split, not a lab number. When you flip through your notes, add one line next to each answer: “Put this advice in my scene — which condition is still unconfirmed?” If you can't write it, let AI ask you one at a time with the template above.

What information the next question needs to carry — see Let AI explain it for my situation. To check whether you actually understand, see The explain-it-clear gate.

If you can write the next question, you've taken the thinking back.