Curiosity is the pull you feel toward a question you want answered. Tom puts the goal simply: “We want to use our natural curiosity, and we want AI to extend it.” An AI chat that has seen your recent conversations knows something about what you wonder about and how you work. Asked well, it points you to what sits just past the edge of what you know, and to the assumptions that are hard to see from the inside.
This guide covers both kinds of edge. The first is curiosity: something new you’d love to learn. The second is blind spots: habits or assumptions that quietly work against your goals. Every prompt keeps the judgment with you, since the whole point is a mind that grows stronger while the AI helps.
The Most Rewarding Questions Are Just Past What You Already Know
The mistake
Using AI only for assigned work, and never for the questions you actually wonder about.
The fix
Ask the AI what you might find fascinating based on your recent conversations, then follow the question that sits just past what you already know.
Your questions to an AI are a record of your curiosity. If you’ve asked about Saturn’s rings and beginner telescopes, the chat has a decent guess about what lights you up. Asking it what you’d find fascinating turns that record into a steady supply of new things to learn. The answer might be an event coming up soon, like a meteor shower in the next few weeks, a surprising fact, or a misunderstanding that most people share.
In 1994 the economist and psychologist George Loewenstein described curiosity as a response to an “information gap,” the feeling of noticing that something you want to know is missing from what you know. One idea in his account is that knowing a little about a topic makes its gaps easier to notice, which helps explain why curiosity often grows as you learn. The most rewarding question to ask sits right at the edge of what you know. A question you already know the answer to teaches little, and one far past your knowledge is hard to hold on to.
The question just past what Hannah knows
Hannah already knows
- A meteor is space debris burning up in the atmosphere.
- Meteor showers return on about the same dates every year.
Which question sits right at the edge of what Hannah knows?
When the AI brings you something new, ask it to question you before it explains. Saying what you already think, even when it turns out wrong, gives the explanation something to correct, and it turns reading an answer into checking your own idea.
Based on our recent conversations, what's something you think I'd find fascinating that I probably don't know yet? It could be something happening soon, a surprising fact, or a common misunderstanding about a topic I've asked about. Before you explain it, ask me what I already think. Let me choose which thread to follow next.
Use this when chat memory is on and you want the AI to build on what you already wonder about.
Here are a few things I've been wondering about lately: [three questions or a short journal entry]. Based on these, what's something related that I'd probably find fascinating and don't know yet? Ask me what I already think before you explain it. Let me choose which thread to follow next.
Use this when memory is off, or in a new app. Paste in three things you've wondered about lately.
I'm learning about [topic]. Here's what I understand so far: [your explanation in a few sentences]. Give me one question right at the edge of what I understand, hard enough to stretch me and close enough that I can reason my way to an answer. Don't answer it for me; let me try first and tell you my thinking.
Use this when a topic has caught your interest and you want a question that stretches you without losing you.
What the AI Remembers From Your Past Chats
The mistake
Sharing everything with an AI that remembers, without checking what it stores or how to turn memory off.
The fix
Check your app's memory settings, use a temporary chat for anything you'd rather keep out of memory, and leave out details you wouldn't want stored.
The curiosity and blind-spot prompts work best when the AI can draw on your past conversations. ChatGPT, Claude, and Gemini each have a version of chat memory, and each one lets you turn it off. Memory makes these questions far more useful, and it also means your chats are stored and reused, so decide on purpose whether to keep it on.
ChatGPT
What it remembers“Reference chat history,” part of memory since April 2025, lets it draw on past chats.
Where to turn it offSettings, then Personalization, then Memory
How to keep one chat outA Temporary Chat stays out of memory.
Claude
What it remembers“Search and reference chats,” on paid plans, looks through past chats when you ask it to.
Where to turn it offSettings, in the memory controls
How to keep one chat outDeleting a chat keeps it out of later searches.
Gemini
What it remembers“Personal context,” added in August 2025, remembers key details from past chats.
Where to turn it offSettings, then Personal context, then Your past chats with Gemini
How to keep one chat outA Temporary Chat isn't used to personalize later replies.
Menu names change often. If a setting has moved, search your app’s help pages for “memory.” Leave out anything you’d keep from a stranger, such as your address, your passwords, or other people’s private information. With memory off, paste in a few recent questions or a journal entry and use the second curiosity prompt above.
What Might You Be Missing About Yourself?
The mistake
Taking an AI's list of your blind spots as a verdict, or waving it away.
The fix
Ask for blind spots with your best intentions assumed and evidence from your own conversations, then treat each one as a guess to test.
A “blind spot” is something about your own thinking that’s hard to see from the inside, such as an assumption you’ve never questioned or a habit you’ve stopped noticing. Tom asks for them directly: where might I be missing something, or assuming something that isn’t helping me move toward my goals as much as I hope? Asking the AI to assume your best intentions keeps the conversation kind and practical, focused on small misalignments worth adjusting.
AI models also have a known pull toward telling people what they want to hear, called “sycophancy.” In April 2025, OpenAI rolled back an update to ChatGPT’s GPT-4o model after it became, in the company’s words, “overly flattering or agreeable.” Asking for evidence from your own chats is a strong guard against an answer that only flatters you.
Which blind-spot answer gives Hannah something to check?
Hannah asksWhere might I have blind spots in my assumptions about my goal of studying astrophysics? Assume my best intentions.
Hannah asked for her blind spots.Which reply is most useful to her?
Even a specific answer is a guess. The AI only sees what you’ve typed, which is a small slice of your life. Treat each blind spot as a hypothesis, something to test before you believe it. The next section turns one into a one-week experiment.
Based on what you know about me from our conversations, where might I have blind spots in my thinking or my assumptions? Where might I be missing something, or assuming something that isn't helping me move toward my goals as much as I hope? Assume my best intentions, and assume I'm just not aware of it yet. For each one, point to what I said that made you think so. Don't tell me what to do about it; let me decide.
Use this with chat memory on, after you've talked with the AI about your goals for a while.
Be candid with me, the way a good coach would, and skip the compliments. Which of the points you just made are you least sure about? For the one you're most sure about, what would I notice in my own week if you're right? Don't soften it, and leave the decision about what to change to me.
Use this as a follow-up when a reply sounds like praise, or so general it could fit anyone.
These prompts are for study habits, goals, and ways of thinking. For anything heavier, like feeling low for a long stretch, talk with a person you trust, such as a parent, a school counselor, or a doctor.
A Week-Long Test Shows Whether a Habit Is Hurting Your Goal
The mistake
Assuming your everyday habits already serve your goal, because the goal matters so much to you.
The fix
Ask where a habit might be working against the goal, then change one thing for a week and read what happened.
Caring about a goal and moving toward it are two separate things, and the gap between them usually hides in ordinary habits. Hannah wanted to study astrophysics and kept asking for summaries of the physics she needed to learn. Her goal was fine. One habit had drifted out of line with it. A small change, tracked for one week, tells you more about a habit than any amount of arguing about it.
Testing a guess about yourself with a week-long change
- The guessI ask the AI to summarize readings before I've read them.
- The change to tryRead each section first, then ask the AI to quiz me on it.
- What to watchWhich quiz questions I miss, and what kind they are.
Your week
Keep the test small: one change, one thing to watch, and seven days. On day seven, read your notes the way a scientist reads data. If the guess held up, keep the change. If the week showed something else, that’s worth knowing too, and you learned it from your own evidence, which makes it far easier to act on than advice.
One of my goals is [your goal]. Here's what a normal week looks like for me: [your week, including when you study and how you use AI]. Where might one of my habits be working against that goal? Give me your best guess and the evidence behind it. Don't make a plan for me; I'll design my own test.
Use this without memory, or when you want the AI to look at one goal and one ordinary week.
I want to test this guess about myself for one week: [the blind spot]. Help me turn it into a small experiment by asking me questions: what one change I'll make, what I'll notice each day, and how I'll know on day seven whether the guess was right. Don't design the experiment for me; keep asking until I've designed it myself.
Use this after a blind-spot conversation, to design a small experiment instead of arguing with the guess.
A week ago I started testing this change: [the change]. Here's what happened each day: [your notes]. Ask me what I noticed, what surprised me, and whether I want to keep the change, adjust it, or drop it. Don't make the decision for me.
Use this at the end of the week, with your daily notes pasted in.
Does Your Request Build Your Skills or Replace Them?
The mistake
Asking AI to do whole tasks from start to finish, until the skills behind them fade.
The fix
Ask AI to question you, quiz you, and show you what's at the edge of your understanding, and keep the reading, solving, and deciding that build the skill for yourself.
Tom’s view is that AI, like any powerful tool, “can be used to create dependency, or it can be used to empower you and make you stronger and smarter and more resilient and more capable.” What decides it is the request. A request that has the AI pull more out of you builds strength, and a request that has it write, solve, or decide for you builds dependence.
Does this request hand the skill over, or build it?
“Write the conclusion for my lab report.”
“Quiz me on this lab before I write my conclusion.”
“Summarize this chapter so I don't have to read it.”
“I read the chapter. Ask me three questions that test whether I understood it.”
“Tell me where my argument is weakest, and let me fix it.”
“Pick which extracurricular I should drop.”
Early research points the same way. In a 2025 survey of 319 knowledge workers by researchers at Microsoft and Carnegie Mellon, people with more confidence in the AI reported doing less critical thinking, and people with more confidence in themselves reported doing more. A small 2025 study from the MIT Media Lab, posted as a preprint, had 54 people write essays with an AI chat, a search engine, or no tools at all. The AI group showed the weakest brain connectivity on EEG, and its members had trouble quoting their own essays afterward. Both studies are early and limited, and both fit one practical rule: keep the part of the work that builds the skill for yourself.
Train yourself to ask the AI to pull more out of you: more questions, more practice, more of what sits right at the edge of your understanding. Used that way, every conversation leaves you a little more capable than it found you.
Look back over how I've used you in our recent conversations. Which requests had you do work that I should be learning to do myself, and which ones made me think harder? Point to specific requests. Don't redo any of that work for me; let me decide what to change.
Use this once a month to look at how you've been using AI, with chat memory on.
Questions to ask yourself
- Did I ask the AI about something I'm actually curious about this week?
- When the AI named a blind spot, did it point to something I really said?
- Am I testing one guess about myself right now, with a note for each day?
- Which of my requests this week made me think harder, and which did the thinking for me?
Sources
- 01George Loewenstein, Psychological Bulletin, 1994The Psychology of Curiosity: A Review and ReinterpretationProposes that curiosity arises when attention falls on a gap in what a person knows or understands.
- 02NASA ScienceGeminidsThe Geminids peak around December 13 and 14, come from the asteroid 3200 Phaethon instead of a comet, and need no telescope or binoculars.
- 03OpenAI, April 2025Sycophancy in GPT-4o: What Happened and What We're Doing About ItOpenAI rolled back a GPT-4o update that had made ChatGPT overly flattering or agreeable.
- 04Lee and others, CHI 2025 (Microsoft Research and Carnegie Mellon)The Impact of Generative AI on Critical ThinkingA survey of 319 knowledge workers: more confidence in the AI went with less critical thinking, and more self-confidence went with more.
- 05Kosmyna and others, MIT Media Lab, 2025 preprintYour Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing TaskA small study of 54 participants writing essays with an AI chat, a search engine, or no tools. The AI group showed the weakest brain connectivity and had trouble quoting their own essays.