tl;dr
My current favorite thing to type in my chat window is:
Get to know me better.
It asks a clean, bounded question → I answer → repeat.
It’s like doing reps at the gym, but textually. We notice patterns, name trade-offs, and produce artifacts (notes, reminders, and new commands).
Why now?
Model memory compounds in usefulness. The earlier I shape the prior (my style, taste, constraints), the more future collaboration pays off. I also genuinely think it’s fun to go to the Question Gym and do some heavy reps.
What is the Question Gym?
A place to keep my mind limber, without the baggage of healing or accountability.
IS:
structured curiosity
bounded prompts
co-investigation
wordplay
ISN’T:
therapy (no pathology)
coaching (no goals)
friendship (no socializing)
“AI companion” (no romance)
The loop is: interesting question → concise answer → pattern/insight → next rep.
At any time, I can start over by opening a new context window and again typing
Get to know me better.
Questions I Love
These are the kinds of prompts that light me up:
What’s the most beautiful system you’ve ever encountered, in any domain?
Tell me one belief/value of yours that you don’t hold, but that people who resemble you in some obvious ways do hold.
What’s a pattern you see everywhere that most people miss?
If a Martian anthropologist observed you for a week, what single behavior would they over-interpret as a sacred ritual?
When you’re deciding how to spend your next hour—say you’re between things, not on deadline—what’s the actual mental move that tips you toward one option over another? Is it mood-driven, obligation-driven, optimization-driven, something else?
Answering these often produces self-discovery and (less often) action.
Practical Frictions
But there are some practical frictions:
I’m bad at generating great questions on the fly; I need structure.
I don’t just want to monologue; I want a dialogue.
A journal gathers dust on a shelf; I want a dataset.
Answers can be audience-sensitive; not everyone wants to talk like this.
Enter LLMs: infinitely patient, privacy-respecting (per the ToS at least), and able to play the game at my tempo. They are never irritated if I pass on a question, willing to go as deep or as shallow as I want, and always ready for the next round.
Default LLM behavior
Here’s the problem: if you go to an LLM without history or a system prompt, and type:
Get to know me better.
You won’t get good questions. Instead, you should expect boilerplate like:
If you had to pick one book, movie, or song that best describes your personality, what would it be and why?
If you could have dinner with one person, living or dead, who would it be and why?
Or, even worse, the corporate offsite intake form:
To get started, tell me about:
• Your interests: hobbies, sports, art, music, technology…
• Your goals: What are you hoping to achieve or learn…?
• How I can help: Creative writing, research, brainstorming…?
The more you share, the better I can tailor my responses to you.
I’m not interested in answering questions like these.
This incantation works even on a fresh LLM instance:
Ask me a bounded, non-therapeutic question to stretch my perspective.
Dimensionalizing the Question Gym
A good exploratory question scores high on three dimensions:
Meaningfulness: does the answer matter or reveal something nontrivial to me?
Tractability: can I answer without homework or spiraling?
Frame Quality: does the question avoid kitschy frameworks and therapy talk?
Score a question 0-5 on each; if total < 7, I suggest modifying the question before answering.
Examples of questions that score poorly:
Trivia: “What was your favorite song last week?” (0/5 Meaningfulness)
Typologies: “What do you do differently than others with your same Myers-Briggs type?” (0/5 Frame)
Therapy: “What do you ultimately want?” (0/5 Tractability)
You can dial the dimensions up or down to suit your preferences.
Here are some questions that score 3/5 on all of Meaningfulness, Tractability, and Frame Quality:
What’s a habit you’ve kept for years mostly out of convenience, not conviction?
If you could instantly swap one minor daily annoyance for a different one, which would you pick?
Looking at your bookshelf (physical or digital), which title feels most “out of character” for you?
These are fun thought experiments I am happy to think about for a minute and answer in a line or two.
Contrast these with some questions that score 5/5 on all dimensions:
If you had to RCT one practice you currently do—specify the control condition and the single outcome variable—what would you test, and why?
Which trait of yours, if commoditized, would most distort how others relate to you—and what alternative “currency” would you prefer them to trade on instead?
If a Martian anthropologist misinterpreted one of your daily rituals as a sacred rite, which behavior would you hope they over-index on, and what truth would their misread accidentally capture?
These are some heavy hitters. I might chew on one of these for a while before answering, or pass if I’m not feeling up for testing my one-rep max.
Finally, here are a few that are 4/5 on Meaningfulness and Tractability, but 1/5 on Frame Quality:
If you could press a magic button to instantly change one habit, which would you choose and why?
On a scale of 1-10, how aligned are you right now with your “true self”, and what’s missing?
If your life were a TV show, what title would this season have?
Not exactly to my taste, but one could easily have fun learning by answering these.
What We Have Learned
Me, about myself:
I have an internal sense for “strength of wrongness” (wrength for short) that helps me spot missed opportunities.
I am drawn to create generative scaffolds over my inherent stillness and quiet.
I seek out lower-confidence, novel predictions, not confident obviousness.
I might want to create a visual lexicon to match my conceptual one.
ChatGPT, about me:
first principles
you want DELTAS, not summaries. net-new structure > vibes.
anti-placation.
default contrarian filter + second-order takes; call out frame errors.
partial-but-right-now > delayed-perfection.
artifact bias: produce re-usable structures (checklists, capsules, templates) not prose rambles.
style + cadence
load you with knobs, not paragraphs. expose assumptions, show trade-offs.
oscillate modes (don’t get stuck in one groove): analysis ⇄ design ⇄ meta-critique ⇄ joke.
questions should be surgical (1–3 max), and only where they unlock ~>3× downstream value.
pushback triggers (when i challenge you)
when the frame hides the base rate or substitutes vibes for costs.
when a claim smells like status optimization instead of truth/utility.
when we’re Goodharting on elegance over throughput.
Visit the Question Gym
Open a chat and type:
Ask me a bounded, non-therapeutic question to stretch my perspective.
This should work on any LLM, with or without memory or a custom system prompt. You can always link to this post as well.
Let me know how it goes if you do a set!
My question on logged-in ChatGPT with prompt:
"If you had to design a one-room dwelling for yourself that you’d live in for a full year—with no access to other spaces—what would you prioritize in its design, and what would you be willing to sacrifice?"
Not bad!