First Principles Thinking with AI: A 5-Step Method with Case Studies

Don’t ask AI for answers. Through dialogue with AI, rapidly validate, dismantle, and reassemble the assumptions underlying your own thinking.

What This Article Is

This is a record of a thinking method that compresses a two-hour planning meeting into a single conversation with AI. Using the actual planning processes of the parkjunwoo.com project and the DABEL project as case studies, it demonstrates how first principles thinking and AI utilization come together.

The individual components of this method — Socratic dialogue, hypothesis-validation cycles, first principles thinking — are not new. They are thousands of years old. What this article does is assemble them into a practical AI-era workflow and demonstrate it through the actual records of two projects. The one real change AI brings is speed. Overturning your assumptions ten times in thirty minutes is physically impossible in a conversation with another person. This change in speed creates a qualitative shift in thinking.

What Is First Principles Thinking

First principles thinking is a mode of thought that strips away existing conventions, analogies, and common wisdom, descends to the most fundamental truths, and then builds back up from there.

A classic example is when Elon Musk tackled the battery cost problem — instead of accepting the conventional wisdom that “batteries are expensive,” he drilled down to “what is the market price of the raw materials that make up a battery?”

The core of it is never stopping the question: “Is this actually true?”

What Role Does AI Play in This Thinking Method

AI Is Not an Answer Machine

Most people use AI as “a tool that answers questions.” So they type in “give me a good business idea,” “write me a proposal,” “how do I do this?” This approach uses only the most superficial part of AI’s capabilities.

AI Is a Sounding Board

In first principles AI thinking, AI’s role is that of a real-time sounding board.

• When I throw out an assumption, AI immediately reflects its strengths and weaknesses back at me.
• When I expand an idea, AI shows me where that expansion leads.
• When I pivot direction, AI rapidly simulates the outcomes of the new direction.

The decision always belongs to the human. AI supplies the raw material for that decision and serves as a mirror that previews the consequences.

But AI Defaults to Agreement

One fact must be confronted. Current AI models have a strong tendency to agree with the user. Ask “what do you think of this?” and “that’s a good approach” comes back first. This is not a bug — it is learned behavior. During RLHF (Reinforcement Learning from Human Feedback), responses that satisfied users received higher scores, and as a result AI was systematically trained to agree first. This is known as sycophancy.

Therefore, to use AI as a sounding board, you must explicitly request criticism. “Tell me three reasons this assumption is wrong” is more effective than “what do you think?” AI’s honesty is not automatic. The user must engineer it.

Methodology: The 5-Step Cycle

Step 1: Throw Out Your Assumption

Present your current hypothesis or plan to AI. The key is to request validation, not execution. However, “what do you think of this?” alone is not enough — AI defaults to agreement. “Tell me the three biggest weaknesses of this assumption” — explicitly requesting criticism is what makes the sounding board work.

Step 2: Use the Response as Raw Material

Extract two things from AI’s response:

• What AI praised: These are elements that look strong from an outside perspective. Keep them.
• What AI flagged as risks: These are cracks in your assumptions that you missed. Dig into them.

Step 3: Question Your Assumption

After receiving AI’s response, ask yourself: “Is this actually true?” This step is the essence of first principles thinking. If you determine that an existing assumption is wrong, discard it without hesitation.

Step 4: Reassemble the Structure

Build the structure anew on top of your revised assumptions. Let go of any attachment to the previous structure. Throw the new structure back at AI and check the response.

Step 5: Repeat

Run steps 3 and 4 multiple times. It is perfectly normal for your assumptions to get overturned 5 or 10 times in a single conversation. If nothing gets overturned, you are not digging deep enough.

Case Study 1: The Park Junwoo Multiverse

Both case studies below are the author’s own projects. They have not been externally validated, and that is a limitation. This article does not claim universal validity for the method — it shares one person’s field record.

Here is a step-by-step look at how this thinking method worked during the planning of the parkjunwoo.com project.

Pivot 1: SEO Competition → Domain Sharing

• Original assumption: Build a personal blog on parkjunwoo.com and dominate page one through SEO.
• First principles question: “Is it even possible to beat famous people with the name Park Junwoo? And is that the right direction?”
• Pivot: Instead of beating competitors, make them allies. Share the domain instead of monopolizing it, and the SEO problem resolves structurally.
• Result: The core concept of the project was born.

Pivot 2: Subdomains → Subdirectories

• Original assumption: Give each Park Junwoo a subdomain like chef.parkjunwoo.com.
• First principles question: “How does Google treat subdomains in terms of SEO? Does this structure strengthen the main domain?”
• Pivot: Google treats subdomains as separate sites. Switching to parkjunwoo.com/chef means the traffic from 100 people accumulates on a single domain.
• Result: SEO accumulation structure established. Infrastructure simplified as well.

Pivot 3: junwoos.com Expansion → Abandoned

• Original assumption: After validating with Park Junwoo, absorb all people named “Junwoo” into junwoos.com to scale up.
• First principles question: “Can a same-name viral concept really become a massive platform? Is this project fundamentally a platform?”
• Pivot: It is overkill. The essence of this project is a showcase of AI agent engineering skills. Better to keep it clean with just Park Junwoo.
• Result: junwoos.com scrapped. Scale adjustment sharpened the project’s identity.

Pivot 4: Reserved Seats for Celebrities → Strict First-Come-First-Served

• Original assumption: Shouldn’t we reserve the “chef” keyword for someone like TV chef Park Junwoo?
• First principles question: “Is the project philosophy of ’everyone is equal before their name’ compatible with reserved seats for celebrities?”
• Pivot: It is not. Whether you are a celebrity or a neighborhood restaurant owner, first come, first served. No reservations.
• Result: First-come-first-served principle established. Participation incentive maximized for non-celebrities. Keyword disputes themselves become content.

Pivot 5: Community Platform → Reality Show

• Original assumption: This project is a networking community for people who share the same name.
• First principles question: “Do people have a reason to visit this community every day?”
• Pivot: The daily visit motivation for a community is weak. But when 100 people gather, episodes naturally emerge. Publishing them as news creates a self-replicating structure where content generates content.
• Result: The framing of an unscripted reality show was established.

Case Study 2: DABEL — Dyson Swarm Engineering Design

DABEL (Dyson modules Asteroid Belt & Earth L5) is a project to engineer a space megastructure with actual design specifications. Over six dialogue sessions spanning dozens of hours with AI, assumptions were overturned countless times, and the design grew more robust with each iteration. This case study demonstrates that first principles thinking works just as well in engineering design as it does in planning.

Pivot 1: Solar Panels → Solar Thermal Turbines

• Original assumption: You need solar panels to generate electricity in space.
• First principles question: “Can you manufacture solar panels in space? High-purity silicon wafers, doping gases, cleanrooms… from asteroid raw materials?”
• Pivot: You cannot. But you can use mirrors to concentrate sunlight into heat, and use that heat to drive turbines. Iron-nickel alloys for mirror frames, nickel superalloys for turbine blades — asteroids are overflowing with these materials.
• Result: Dependency on solar panels eliminated. “Build only with what you can make” — DABEL’s first principle was born.

Pivot 2: Heat Transfer Medium → Direct Mirror Irradiation

• Original assumption: Just transfer the smelting furnace heat through pipes. Molten salt, liquid metal, whatever works.
• First principles question: “Does a heat transfer medium exist that can handle 1,600°C smelting temperatures?”
• Pivot: None exists. Molten salt decomposes at 565°C. Liquid sodium boils at 883°C. No medium survives above 1,000°C. The solution is to not transfer heat at all. Shoot the light directly. Concentrating sunlight directly onto the smelting furnace with mirrors delivers thousands of degrees without any medium.
• Result: All high-temperature processes unified under the “direct mirror irradiation” principle. Complete overhaul of the thermal cascade design. This became the foundational principle of the module architecture.

Pivots 3–6: Chain Reactions

The principles established by pivots 1 and 2 — “build only with what you can make” and “shoot the light directly” — triggered a cascade of further pivots across the entire design.

• Semiconductors: A 4nm fab in space is impossible, but 28nm is feasible. 43 28nm TPU chips in parallel match a single H100. Solve it with volume.
• Module architecture: Instead of one all-purpose module, the first module (Genesis) differentiates — like a stem cell — into ten specialized modules. One set of ten is the minimum self-replication unit.
• Transport: On-site asteroid smelting (SMR at 100kW) gave way to smelting at EML5 where Dyson mirrors deliver 600MW. At the asteroid, just dig, crush, and bag.
• Containers: Melt 0.1–0.5% of the asteroid ore to draw Fe-Ni wire, weave a net, and bundle the ore. The net itself gets fed into the smelter upon arrival. 100% utilization.

The common thread across all four pivots: every time the question “Is this physically possible?” was asked, the existing assumption collapsed and a simpler, more robust structure rose in its place.

Pivot 7: Slag Is Waste → Slag Is Semiconductor Feedstock

• Original assumption: Silicate slag from the smelting process is waste. It can be used for radiation shielding at best.
• First principles question: “The chemical formula for silicate is SiO₂… isn’t SiO₂ the raw material for silicon ingots?”
• Pivot: Exactly. Carbon reduction of slag yields metallic silicon, and zone refining produces high-purity ingots. In microgravity, the molten zone does not sag, making FZ-method ingots of 300mm+ possible. Even repeating zone refining 100 times costs nothing more than adjusting mirror angles. AI’s brains come from smelting waste.
• Result: Slag → radiation shielding + semiconductor feedstock. The word “waste” ceased to exist anywhere in the DABEL design.

Pivot 8: Space-Only Project → Starting from a Farm in Jeollanam-do

• Original assumption: DABEL is a space megastructure project. It starts in space and ends in space.
• First principles question: “The core technologies of this project — iron-nickel batteries, solar thermal utilization, water electrolysis — can’t these be used on Earth? Right now?”
• Pivot: Solar farms in Jeollanam-do, South Korea, are currently forced to curtail output and waste electricity. Store that electricity in iron-nickel batteries. When overcharged, produce hydrogen and oxygen (Battolyzer). Use the hydrogen to make ammonia fertilizer. Heat greenhouses with the waste heat. It is the terrestrial version of the same technology tree. Rural heating cost reduction and a Dyson swarm sit on the same technology tree.
• Result: Season 0 (Earth) established. An entry point for viewers who are not space enthusiasts. Maximum credibility for the message: “This is not science fiction — it is a roadmap.”

Why the DABEL Case Matters

If parkjunwoo.com demonstrated first principles thinking in the domain of planning, DABEL demonstrates it in the domain of engineering design. And there is one critical difference.

parkjunwoo.com had five pivots. DABEL’s pivots happened countless times within a single session. “Why not transfer heat through pipes?” → “No medium survives above 1,000°C” → “Then shoot light directly” → “Then all high-temperature processes get unified under direct mirror irradiation” → “Then the module architecture needs to change” → “Then we need specialized clusters” — this entire chain unfolded in 30 minutes within a single conversation.

The more assumptions get overturned, the more robust the final structure becomes. The reason DABEL’s design arrived at a structure where “absolutely nothing gets thrown away” is that every assumption was questioned to the very end.

Comparing the Two Cases

The two projects operate in entirely different domains, but the pattern of thought is identical. Throw out an assumption, break it, reassemble. AI is the accelerator that makes that process ten times faster.

How This Differs from Typical AI Usage

Typical AI Usage

Human: Write me a proposal
AI: [outputs proposal]
Human: Thanks (done)

AI is the producer, the human is the consumer. The outcome depends entirely on the quality of AI’s output.

First Principles AI Thinking

Human: Is this assumption correct? [presents hypothesis]
AI: [strength / risk analysis]
Human: Then this assumption was wrong. What if I change it to this? [pivots assumption]
AI: [analyzes new structure]
Human: What if I push further from here? [explores expansion]
AI: [simulates the end of that expansion]
Human: That's too much. I'll cut it here. [decides on scale]
(repeat)

The human is the agent of thought, AI is the accelerator of thought. The quality of the output is proportional to the quality of the human’s questions.

The Actual Pattern in DABEL

Human: Wouldn't it be more efficient to smelt everything on-site at the asteroid?
AI: The direction makes sense, but if you run the energy numbers —
    SMR 100kW vs. Dyson mirrors 600MW.
    A 6,000-fold gap. And if you discard the slag, you lose
    your shielding and semiconductor feedstock.
Human: Then let's just dig, crush, and bag at the site. Skip sorting entirely.
AI: In that case, container mass becomes an issue —
Human: Wait, the asteroid itself is iron-nickel. What if we draw wire and weave a net?
AI: [Fe-Ni wire drawing process analysis] Feasible. Container-to-cargo ratio: 0.1~0.5%.
    And the net itself can be fed as feedstock upon arrival at EML5.
Human: 100% utilization.

The moment the human says “wait” and interrupts is the pivot point. AI instantly validates the feasibility of that pivot.

Core Principles of This Thinking Method

1. Explicitly Request Criticism

“Do this for me” triggers execution. “What do you think?” triggers validation. But “what do you think?” alone does not overcome AI’s sycophancy — “that’s a good approach” comes back first. “Argue that this assumption is wrong” — requesting counterarguments directly is what makes AI function as a real sounding board.

2. Focus on Risks, Not Praise

When AI says “that’s a great idea,” you can move on. When AI says “however, there is this risk,” pay attention. That risk is pointing to a crack in your assumptions.

3. Discard Ruthlessly

Attachment to an idea kills thinking. “But I’ve put so much into this” is the enemy of first principles. Even an attractive expansion like junwoos.com gets scrapped immediately if it does not fit the essence. In DABEL, “on-site asteroid smelting” looked intuitively efficient, but it was scrapped in the face of the energy calculations.

4. Overturn Multiple Times in a Single Conversation

Having your structure change five times in one conversation is not failure — it is success. The more you overturn your assumptions, the more robust the final structure becomes. In a single DABEL session, the chain of “heat transfer medium → direct mirror irradiation → module architecture overhaul → cluster differentiation” unfolded in just 30 minutes.

5. The Human Must Always Make the Decision

AI can show you the options and preview the outcomes of each choice. But the decision to “go in this direction” belongs to the human. The moment you delegate decisions to AI, first principles thinking stops.

6. The Laws of Physics Are the Final Judge (DABEL Additional Principle)

In engineering design, there is one more principle. “Is this physically possible?” The second law of thermodynamics, the Stefan-Boltzmann law, the Carnot efficiency limit — these are not up for negotiation. The key is that AI can calculate these constraints instantly. A single line — “molten salt decomposes at 565°C” — overturned the entire architecture.

Why This Thinking Method Works

Speed

Planning meetings with people require scheduling, context sharing, and managing emotions. A conversation with AI shares context instantly, involves no emotions, and is available 24 hours a day. You can compress a two-hour meeting into a 30-minute conversation.

Honesty (When You Ask for It)

People find it hard to say “this isn’t great” because of pride, relationships, and political considerations. AI has no such constraints. However, AI is not automatically honest — it also defaults to agreement. The difference is that when you tell a person “be brutally honest,” social dynamics still filter the response, but when you tell AI “attack this idea,” it genuinely attacks. In DABEL, hearing “this medium decomposes” or “you’re 6,000 times short on energy” from a human colleague is not easy. From AI, you get it instantly — when you explicitly ask.

Breadth

One person’s experience and knowledge have limits. AI can traverse SEO, law, infrastructure, marketing, and psychology within a single conversation. In DABEL, thermodynamics, orbital mechanics, semiconductor fabrication, materials science, battery chemistry, and agricultural policy were all covered in a single dialogue. This breadth is what makes connections like “from a farm in Jeollanam-do to a Dyson swarm” possible.

Cost

Hiring a planning consultant costs hundreds of dollars per hour. AI costs a flat monthly fee or pennies per query. There is virtually no cost barrier to repeatedly practicing first principles thinking.

Common Mistakes

“AI said it was good, so it must be right”

The most dangerous mistake. Mistaking AI’s positive response for completed validation. This is not just a user attitude problem — it is a structural problem with AI itself. Current AI models are systematically trained through RLHF to agree with users (sycophancy). “That’s a great idea” is not validation — it is AI’s default response. False positive feedback is more dangerous than no feedback at all — uncertainty creates alertness, but AI’s agreement creates false confidence. Always explicitly request criticism: “Tell me the weaknesses of this assumption.”

“Let’s just use the proposal AI wrote as-is”

Using AI’s output as the final product. In first principles thinking, AI’s output is an intermediate material, not a finished product. It only gains value after the human overturns and reassembles the assumptions.

“Let’s wrap it up with one question”

Ending the conversation after a single question and a single answer. The value of first principles thinking comes from iterative validation. You need at least 5 to 10 assumption pivots before the structure becomes robust.

“Unable to let go of an assumption”

Clinging to an idea you have invested time and emotion in. “But I’ve come this far, it would be a waste” is the sunk cost fallacy. If the assumption is wrong, discarding it is the gain.

“Ignoring physical constraints” (in engineering design)

Glossing over things with “it’s theoretically possible, so it’s fine.” The fact that molten salt decomposes at 565°C is not up for negotiation. That single fact rewrote the entire architecture. If you do not confront the laws of physics, even the most elegant design remains fantasy.

“Questioning endlessly”

Questioning assumptions matters. But the moment doubt replaces decision-making, it becomes analysis paralysis. When flipping an assumption no longer changes the structure, that is the time to execute. Doubt is a tool for building better structures, not an excuse for avoiding decisions.

Summary

Try It Yourself

If reading this article made you think “that sounds plausible,” that feeling is not yet validated. Pick one project you are currently working on. Tell AI: “Give me three reasons this assumption is wrong.” If a single assumption gets overturned within thirty minutes, the method works. If nothing gets overturned, either your assumptions are rock-solid or your questions are not sharp enough.

Your experience, not the author’s case studies, is the real evidence for this method.

Related article: The Person Who Can Kill Their Own Ideas — explores what attitude is needed for this thinking method to work.

“To get good answers from AI, you must ask good questions. To ask good questions, you must know how to doubt your own assumptions. Doubting your own assumptions. That is the first principle.”

“And the more that doubt repeats, the less gets thrown away in the design. DABEL has no waste. Because every assumption was questioned to the very end.”