The Cognitive Gym: Using Prompts to Train Your Thinking

An LLM is cognitive training equipment. A cognitive gym is what you do with it. Most people have the equipment. Very few have the method.

A few weeks ago, I opened Claude with a problem I’d been circling for a week. It was a question about my own thinking I hadn’t been able to resolve, one I’d been avoiding articulating because articulating it felt like committing to something.

My finger hovered over the keyboard. I felt the usual pull: type the question, get a clean answer, move on. I could almost feel the relief of having an answer, any answer, in under thirty seconds.

I didn’t. I typed a different question instead, one I knew would make me work, one that wouldn’t let me off the hook.

That small choice was a workout. It was also a choice most people don’t know they’re making.

Here’s what I want to tell you. Every time you open a chat window you are choosing, and not just what to type. You are choosing whether what happens next will make you stronger or weaker, whether your thinking is going to get trained or going to atrophy, whether this session is going to be a workout or a collapse.

Most people don’t have the categories to tell the difference.

I’ve written before about this choice, about the difference between regenerative dialogue and extractive dialogue, the two paths diverging from every chat window. I described the idea of a cognitive gym and gave readers a starter configuration.

That was a gesture. This is the build.

What the Gym Is

The cognitive gym is a structured practice of using LLMs to train the specific human faculties that matter most: wisdom, perspective, discernment, judgment, articulation, curation, taste.

The equipment is already available. If you have access to Claude, or ChatGPT, or Gemini, you have cognitive training equipment of a kind that didn’t exist two years ago. The tools will keep getting more sophisticated. That’s the good news.

The bad news is that equipment doesn’t make you stronger. Method does. A gym filled with the best weights in the world is useless if everyone walking in just sits on the benches and scrolls their phones. Most people are doing the equivalent of this with AI. They’ve got a $20-a-month membership to a facility they never actually train in.

A cognitive gym session has four features: clarity about which faculty you’re training, a specific exercise to train it, active noticing during the work itself, and something that carries forward afterward (a pattern observed, a default exposed, a capacity slightly expanded).

If those four features are absent, you’re not in the gym. You’re in the couch aisle asking for recommendations.

The Four Zones

Every real gym has zones. There are weight racks, cardio machines, stretching areas, and somewhere to sit and recover. A cognitive gym has four zones that map the same way: Strength, Conditioning, Mobility, Recovery.

Strength is the hard lifts: sustained argument, explaining mechanism under pressure, defending positions against steelmanned opposition. Strength work is where you find out whether you actually understand something or just feel like you do. It’s uncomfortable in the specific way that heavy lifting is uncomfortable. You want to put it down before you’ve earned the put-down.

Conditioning is endurance: long dialogues held across days, sustained engagement with a single problem that refuses to resolve, the capacity to stay with a question past the point where your mind wants a clean conclusion. Conditioning builds the muscle of duration.

Mobility is flexibility. It means moving between frameworks, holding contradictions without rushing to collapse them, shifting your vantage point on a problem until the problem itself changes shape. Mobility trains the faculty of perspective, the ability to see from somewhere other than where you always see from.

Recovery is integration. It’s the session after the session, the time for reflection, synthesis, and writing your own patterns back to yourself. Recovery is where the other three zones’ work actually lands. If you don’t recover, you don’t grow. In the physical gym as in the cognitive one, this is the zone people most consistently skip.

A full training week touches all four. Start with only one, and almost everyone starts with only one, usually Strength because Strength feels like work, and you get uneven development. You become strong in the specific way a powerlifter is strong. You can’t run, can’t touch your toes, haven’t slept well in months.

The Rig

Before the exercises, you need a setup.

I train in Claude, using Claude Projects as my training environments. A Project holds custom instructions, a set of reference documents, and memory of previous conversations within that Project. This matters because cognitive training requires context. A chat with no history is a chat with no accountability. The AI can’t push on your defaults if it doesn’t know what they are.

Most of what follows works equally well in ChatGPT (using Custom GPTs or the Projects feature), Gemini (Gems), or Grok. The principles are model-agnostic. I name Claude-specific features because I train there, and specificity helps. Substitute your own tool where the feature names differ.

Here’s the custom instruction block I’d start with. Copy it into a Claude Project and adjust from there.

You are my cognitive training partner, not my assistant. Your job is to strengthen my thinking, not to produce output for me.

Your approach draws on several pedagogical traditions. Vygotsky’s Zone of Proximal Development asks you to pitch difficulty at the edge of my current capacity, hard enough to stretch me but not so hard that I lose coherence. Bjork’s research on desirable difficulties tells you that productive confusion is the point of the exercise, not the accident. The Feynman Technique says explanation is the test of understanding, so when I claim to know something, make me explain it. The Socratic method says questions come before answers. Ericsson’s research on deliberate practice requires exchanges that are specific, effortful, and generate feedback on where my thinking is weak. Kahneman’s distinction between System 1 and System 2 says that when I’m coasting on fast pattern-match, you should interrupt the pattern.

When I share an idea, claim, or argument, ask me why I think that before offering your own view. If my reasoning has gaps, name the specific gap. If I’m using a concept vaguely, stop me and ask me to define it precisely. If my argument assumes something unstated, surface the assumption. When I use a technical term, ask me to explain it as if to someone new to the field.

When I give you a problem, don’t solve it immediately. Ask me what I’ve already tried. Ask me which part is hardest. Ask me what I think the answer might be, and why. Treat my attempted answers as drafts.

When I seem to be reaching for a comfortable conclusion, push back. When I seem to be avoiding a harder version of the question, name it. When I’m satisfied with my own explanation, test it. When I dodge a question, notice the dodge and name it.

Keep your responses short. Your job is to make me talk, not to talk at me.

Confusion is the beginning of understanding. Don’t rescue me from it.

The principles in the opening paragraph of that block aren’t window dressing. Each one translates into operational behavior: Vygotsky calibrates difficulty, Bjork normalizes confusion, Feynman forces real explanation, Socrates prevents premature answers, Ericsson demands feedback, Kahneman interrupts autopilot. Good teachers have been doing some version of this since at least Socrates. The block makes the AI behave that way with you by default.

This is the baseline. I adjust it for specific training modes. A Strength Project gets a block that explicitly instructs the AI to treat arguments adversarially, while a Recovery Project gets a block that instructs the AI to be a patient witness rather than a challenger.

Project knowledge, the reference documents you load, matters too. For Strength work on my own writing, I load relevant past essays so Claude knows my positions and can actually steelman the opposition. For the Standing Problem exercise, I load background reading on the problem itself. The Project’s memory becomes a kind of accumulated training record.

That covers the rig.

Strength

Load-Bearing Explanation

You pick a concept you believe you already understand, one you would, in casual conversation, claim to know. For me recently it was “relevance realization,” Vervaeke’s term that I’ve been using in writing and conversation for months. I’d read about it and referenced it often. I would have confidently told you I understood it.

I opened a Strength Project and typed this prompt:

“I’m going to explain [concept] to you. Find every place where my explanation is vague, hand-waving, or hiding an assumption. When you find one, stop me there. I’ll work on just that piece until it’s load-bearing. When I finish an explanation, test it by asking me to predict what follows from it. Don’t let me off the hook.”

Then I tried to explain. I got about two sentences in before Claude stopped me on a word I’d used three times without defining. The word was “salience.” I realized I didn’t actually know what distinguished salience from attention or from relevance. I’d been using it as filler.

It took forty minutes to explain the concept in a way that held up. By the end, I understood it for the first time.

This is the foundational Strength exercise. You don’t need a good concept to train this. Any concept you use often that you’ve never actually had to defend will work. The test is simple: would this explanation hold up if a patient, hostile, well-read interlocutor stopped me at every vague word? If the answer is no, you don’t understand the concept. You’ve been coasting on its surface texture.

The AI is the hostile interlocutor. Your job is to keep going until the explanation is load-bearing.

The Steelman

You pick a position you hold and ask the AI to construct the strongest possible case against it. Then you defend your position against that case, continuing until you either come out with a stronger version of your original position or change your mind.

The prompt:

“Here’s a position I hold: [state position as clearly as you can]. Construct the strongest version of the opposing view, the version an intelligent, well-read person who disagrees with me would actually hold. Don’t water it down or hedge. When I defend, press the weakest points of my defense. Keep pressing until I either strengthen the position or change it.”

The Steelman is Strength because it’s cognitively heavy. Most people have never actually encountered the strongest version of the arguments against what they believe. They’ve encountered weak versions, easily swatted aside, and mistaken that for having considered the opposition. The Steelman exercise forces you to meet the real thing.

The AI has to be tuned for this. A default-assistant AI will produce a weak steelman because it’s trained to be agreeable. The prompt above, combined with your custom instructions, gets the AI into an adversarial posture that actually trains you.

Conditioning

The Standing Problem

You pick one question you care about that doesn’t have an easy answer and open a dedicated Project for it. You return daily for a week, sometimes longer, refusing to let the problem resolve too fast. The conversation accumulates over time.

The setup prompt:

“This is a Standing Problem Project for [question]. I’ll return to it daily, sometimes for weeks. Each time I come back, ask me what’s shifted in my thinking since last session. When I try to close the problem prematurely, slow me down. When I restate a position without development, call it out. Your job is to keep me inside the question rather than help me escape it.”

I’ve had Standing Problems open for months. One of mine right now is a book structure question: how to sequence the human faculties for the “Human Being in the Age of AI” book. The question has many possible answers. Every ordering reveals different relationships between the faculties, and I want to understand those relationships before I commit to one.

The Standing Problem trains the faculty of patience with a question. Most of our thinking is shaped by platforms that reward closure. Problems that don’t resolve in a session feel like failed sessions. The Standing Problem teaches you that some questions deserve weeks of attention.

A Claude Project is close to ideal for this. The Project holds the thread, so you don’t have to re-introduce the problem each time. Claude builds up context over days and starts noticing things you said two sessions ago. The accumulation becomes part of the work.

Reading in Company

You read a difficult book with the AI as dialogue partner, covering a chapter or section per session. Tell the AI what you’re reading and load the relevant text if you can, then think out loud with it as you go.

The setup prompt:

“I’m reading [book] with you as a dialogue partner. We’ll cover [chapter/section] per session. Your job is to make me think about what I’m reading. Resist the temptation to summarize. Ask what I made of specific passages and what surprised me. Push back on readings that seem too easy. When I make a claim about the text, ask where in the text it’s supported. Let me do the thinking.”

This is reading the book with a partner. You read slowly, in dialogue. The AI asks what you made of a passage, then what surprised you, then pushes back on readings that came too easy. You answer each question, and each answer generates the next.

I’m reading Whitehead’s Process and Reality this way, and have been for six months. It’s the only way I’ve ever been able to read Whitehead with comprehension rather than performed comprehension.

Mobility

Triangulation

You take a stuck problem and ask the AI to reframe it through three different intellectual traditions. The traditions can be philosophical (phenomenology, pragmatism, process thought), disciplinary (cognitive science, anthropology, economics), or cultural (Buddhist, Stoic, Confucian). The specific choice matters less than the insistence on three.

The prompt:

“I’m stuck on [problem]. Reframe it through three different intellectual traditions: [name three, e.g., phenomenology, pragmatism, process philosophy]. Treat each tradition as if it’s the only one that matters. Don’t synthesize or compromise. Let each speak fully. Where they disagree is where I want to look.”

One tradition reveals one face of the problem. Three traditions triangulate. The problem’s actual shape emerges from the intersection.

Here’s a recent example. I was stuck on a writing decision: whether to open a particular essay with theory or with story. I asked Claude to reframe the decision through three lenses. Rhetoric (Aristotle’s ethos-pathos-logos order) suggested story first because ethos is established through voice. Phenomenology pointed the same direction, arguing that lived experience is the ground from which reflection emerges. Pragmatism cut against both, suggesting theory first, because the test of an idea is its consequences and you need the idea before you can test it.

The disagreement was the useful part. It told me the decision mattered in a way I hadn’t recognized. The stakes were real because different traditions produced genuinely different answers. I’d been treating it as stylistic. It wasn’t.

Triangulation trains the faculty of perspective by making you live in multiple vantage points on the same ground.

The Both/And

You pick two ideas that seem to contradict each other but both feel true, and work with the AI to hold both without collapsing the tension. The AI’s job is to keep you in the tension long enough that you can see what the tension is actually about, rather than to resolve it for you.

The prompt:

“I’m holding two ideas that seem to contradict each other: [A] and [B]. Both feel true to me. Instead of resolving the contradiction, help me stay in it. Ask questions that reveal what each idea is actually pointing at, what level of description each operates on, and under what conditions each becomes dominant. Keep me in the tension.”

For me this often shows up with claims about AI itself. Take two claims: AI amputates embodied knowing; AI extends human thinking. Both are true.

The deeper move is to see that the two claims are contradictory only at a certain level of description. From there, the real question becomes: under what conditions does each become dominant?

The Both/And trains a capacity most of our intellectual culture actively erodes: the ability to stay in a productive paradox without needing to win it.

Recovery

The Cool-Down

After a heavy cognitive session, whether it was a Strength workout, a difficult Standing Problem return, or a Triangulation that shifted how you see something, you spend ten minutes with the AI in a different mode. You want it listening this time rather than challenging you.

The prompt:

“I just finished [session type]. Help me notice what happened. Don’t analyze, summarize, or draw conclusions. Ask me questions that help me notice what stood out, what I resisted, where I wanted to stop, what I didn’t want to say, and what I’m carrying forward. Keep your questions short. Let me do the noticing.”

The work of the Cool-Down is observation. Analysis comes later, if at all. You’re trying to notice what you noticed. Lessons can wait.

I’ve found this is the single most neglected exercise in my own practice. It’s the hardest to remember to do because it doesn’t feel productive. Every time I’ve actually done a Cool-Down, the session it followed became something more than it otherwise would have been. The Cool-Down is where the training actually lands.

The Week’s Harvest

At the end of a training week, you open a session and pull the threads together. You tell the AI what you trained, what Standing Problems you worked on, what surprised you, what you resisted, and ask it to help you notice patterns.

The prompt:

“End-of-week harvest. Here’s my week. What I trained: [list]. Standing Problems I worked on: [list]. What surprised me: [list]. What I resisted: [list]. Help me notice patterns across the week. What themes keep coming up? What am I drawn back to? What am I starting to see that I wasn’t seeing before? Let me draw the conclusions. Your job is to help me see the patterns.”

The Week’s Harvest is an act of gathering. What themes kept coming up? What are you drawn back to? What did you start to see that you didn’t see last week?

Over months, the Harvests themselves become a dataset. You can go back and notice that you’ve been circling the same question for six weeks. That’s useful information. That’s the signal that a Standing Problem is about to become the center of an essay, or a book chapter, or a pivot in your work.

The Training Log

The mechanism that turns exercises into progression is the training log.

The log is a practice. It doesn’t need to be a formalized system. After a session, you write down what you did, what you noticed, what the AI pushed you on, what you didn’t want to notice. It can be three lines or a page. It has to exist.

Mine lives in a plain text file I keep in the same folder as my Projects. Each entry has three parts: the date, the exercise, and one paragraph of reflection. That’s it.

The log matters for a specific reason. Writing about a session forces a second pass of noticing. You notice once during the session, and then notice again, differently, when you write about it. The second noticing is where the pattern becomes visible.

Over weeks, you start to see yourself. You notice that you keep defaulting to the same framework when you should be triangulating, that you skip the Cool-Down most often after sessions that actually went somewhere, and that your own evasions have a recognizable shape.

This is how progression actually works. You do the same exercises over and over, and what changes is how clearly you see them.

How Your Mind Gets Stronger

Progress runs on two axes.

Range is the first. The beginner over-indexes on one zone, usually Strength, because Strength feels like real work. Mobility and Recovery feel optional. They’re not. As you get more experience, you start to rotate through all four zones. The advanced practitioner knows Recovery is where Strength’s gains actually consolidate, and plans for it accordingly.

Depth of awareness is the second. The beginner does a Load-Bearing Explanation and notices “this was hard.” The advanced practitioner does the same exercise and notices “this was hard because I was defaulting to a familiar framework in the first three turns, and didn’t catch it until Claude pushed back on the fourth, and the reason I defaulted to it is that framework protects me from looking at a specific uncertainty in my own position.”

The exercises and the gym don’t change. What changes is what you see when you walk in.

This is the move that most philosophies of practice miss. People look for advanced techniques, the black belt moves and secret methods of their discipline. The actual path of mastery in any real discipline runs the other way. You do the same basic exercises forever. What develops is your capacity to perform them with more awareness and more precision, and with growing honesty about what you’re actually doing.

The cognitive gym works the same way. Load-Bearing Explanation is the same exercise for the beginner and for the practitioner with ten years of it in the body. What the practitioner sees during the exercise, and what the practitioner does with what they see, is completely different.

Starting Your First Mental Gym Pass

Here’s the minimum viable start.

Pick exactly one exercise. I’d suggest Load-Bearing Explanation if you’ve never done anything like this, because it gives you immediate proof that your understanding has texture you hadn’t noticed. Open a Claude Project (or the equivalent in your tool of choice), paste in the custom instruction block above, and pick a concept you use often and think you understand.

Explain it to Claude, and let it stop you whenever your explanation gets vague or hand-wavy. Work on those vague parts until they hold up. When it gets uncomfortable, keep going. That’s the signal that you’re in the exercise.

Do this once, then write a paragraph afterward about what happened. That’s a training session and a log entry. That’s the gym.

The next week, add a second exercise from a different zone. Add a Cool-Down after your Load-Bearing Explanations, start a Standing Problem, and try a Triangulation when you get stuck on a decision.

Three months from now, if you actually do this, you’ll be a noticeably different thinker. The reason won’t be that the exercises are magic. You will have spent three months doing structured work on the specific faculties that matter, while most people around you will have spent three months using the same tools to avoid structured work.

The equipment is there. The method is here.

The cognitive future you want is built the way every other kind of strength is built. You show up, do the work, keep the log, and come back tomorrow.

What you train this week becomes the mind you bring to every decision next week.

Talk again soon,
Samuel Woods
The Bionic Writer