Ask an LLM — and go further than the article does
You’re reading an article. The topic pulls you in. But after the conclusion, a question appears that the article didn’t ask — or asked, but not in the direction that interests you.
You can close the tab. You can look for another article that might not exist. Or you can ask a language model — and choose your own direction of exploration.
This isn’t a guide to AI. It’s a demonstration of one specific technique for working with LLMs, based on two articles from AI907. I’m using Claude (Anthropic) as an example — because that’s the model I currently reach for — but the technique described works just as well with ChatGPT, Gemini, Grok, or any other model you have at hand.
One caveat upfront, before I show anything: LLMs make confident mistakes. They give wrong numbers in an expert’s tone. They cite sources they haven’t verified. They generate responses that sound credible, because that’s how they were built. We’ll come back to this.
The responses I quote below are shown in raw form — without iteration or verification. They are demonstration examples, not finished material. Both would require several rounds of refinement before they could be used for anything serious.
Topic one: the regolith shield
The article about the regolith shield ends with the observation that the real challenge of building a lunar base is not the habitat design, but moving hundreds of tons of material. The physics is described. The engineering gap — between knowledge and action — is too.
But for me, a different question appeared after that article. Not “how many tons?” — that would just deepen the same thread. My question was systemic: does placing four habitats with shared walls change the equation — both in terms of regolith mass and heat retention through the lunar night?
The model offered me three other directions to continue. None of them interested me. I chose my own.
Prompt:
“Does placing habitats adjacent wall-to-wall (e.g. 4 together in a 2×2 layout) make sense to minimize the amount of regolith needed and heat retention through the night? How do you prepare the contact surfaces to connect modules after the night?”
Claude’s response (excerpt):
“Yes, and it makes a great deal of sense — but with precisely defined caveats.
Each shared wall is one surface that doesn’t need to be covered with regolith from the outside. With 4 modules in a 2×2 layout, you have 4 shared walls instead of 16 free-standing ones. That realistically reduces the required regolith mass by 20–35% depending on the module shape and required layer thickness.
The thermal effect is actually greater than the radiation effect. A shared wall between two heated modules is actively insulated on both sides — instead of being in contact with the vacuum at -170°C on the dark side. The heating system has to fight a smaller heat-loss surface.
The joint problem is genuinely difficult operationally and is almost absent from popular literature. Aluminum and titanium steel change their linear dimensions by about 0.2–0.5% of their length with a 200–300°C temperature drop. For an 8-meter module, that’s 16–40 mm of movement in each dimension. Three approaches to the joint: a compensating joint (expansion joint), a dry disconnectable joint, or a monolithic system with a unified shield buried under one regolith pile — then the entire system behaves as a single thermally inert object.
The most serious problem is regolith dust being ground deeper into the seal with each thermal movement. Standard approach: soft gold or indium metal seals, labyrinth joint geometry, a spring constantly pressing the joint.”
A few things in this response are worth noting. First — the LLM not only answered the question, but identified an aspect that “is almost absent from popular literature.” I can verify that. Second — it gave specific numbers (20–35% mass reduction, 16–40 mm of movement). These are points that require verification before I use them in any serious context.
Topic two: the space of human life
The article about human living space ends with a philosophical question: if only 0.26% of the planet’s volume is accessible on Earth, perhaps expansion into space is the logical consequence? That’s a good starting point — but I was interested in a different direction.
Prompt:
“If we wanted to double the space available to humans without leaving Earth — what would be technically simplest: going deeper underground, building upward, or air-conditioning the Sahara?”
Claude’s response (excerpt):
“Going underground is technically the most difficult of the three directions, counterintuitively. Below -1 km, temperature rises by ~25–30°C per kilometer of depth. At 3 km you have ~80°C, at 5 km you exceed 100°C. At 3 km depth, lithostatic pressure is ~80 MPa. Every tunnel requires either active support or materials only used in specific industrial applications.
Building upward is technically cheapest — and here there’s a subtle trap. Vertical building doesn’t increase available volume in the sense of the article, it densifies it. If you have 8 billion people in 2 km-high skyscrapers instead of single-family homes, you’re not adding anything to the total volume of the biosphere.
Air-conditioning the Sahara — if the question is ‘what is technically simplest to add space’ — then developing dry deserts is the only option that actually expands the biologically accessible layer, while not fighting basic material physics. Real ‘air-conditioning the Sahara’ isn’t literally cooling 9 million km² — it’s green infrastructure, closing the water cycle. The Sahara at its climate optimum ~8,000 years ago had lakes, rivers, and savanna. This isn’t terraforming, it’s a return to conditions that existed recently on a geological timescale.
There’s also a fourth direction that the question doesn’t mention — one the article suggests directly as a conclusion: orbit and the Moon are in some ways easier than 5 km below Earth’s surface, because at least you’re not fighting the heat of the planet’s interior and not carving tunnels through rock under 80 MPa of pressure.”
I draw attention to the last paragraph. The model itself connected the thread of the question with the topic from the article — and proposed a fourth direction that wasn’t in the question. Is that direction interesting? Yes. Did I choose it? No. I stayed with the Sahara and wanted to know if anyone had already tried to realize such concepts. Another prompt, another iteration.
It’s worth noting the technical context: I attached the source article content to both prompts. That’s why in the response to the question about going underground, the model operates with depths of 1 km and more, and for building upward — with heights around 2 km, because those were the values that appeared in the article. If I were to iterate this thread further, I’d start with depths of tens or hundreds of meters, not a kilometer. The response isn’t wrong — the prompt was insufficiently specified. This is also a practical observation from working with a model: it answers the question you asked, not always the one you had in mind.
How to verify LLM responses
Two methods, both quick.
Method one: a follow-up question about certainty.
When the model gives a specific number or claim that you want to use further — ask directly:
“You said that the regolith mass reduction in a 2×2 layout is 20–35%. Are you certain about this number? Where does it come from?”
The model will either cite a source (which you can check), or say it’s an approximation without specific grounding. Both results are valuable. The second tells you that you can’t cite the number — only treat it as an orientation for further searching.
Method two: ask for a definition instead of a source.
A stronger test than “where did you get that?” is “how do you define that?” The model often operates on memorized approximations whose definitions are inconsistent between responses. The question “how do you define ‘biologically accessible space’?” reveals inconsistencies faster than asking for a source.
I use both methods regularly. Neither replaces checking primary sources when numbers need to carry weight beyond private exploration.
What follows from this
An article has one direction — the one the author chose to describe. A conversation with an LLM has as many directions as you have questions.
A few observations from practice:
An LLM will always propose a direction. It won’t always be your direction. It’s worth having your own question before you open a conversation — because the model will happily fill the void with its own suggestion, and that suggestion can be interesting, but it’s rarely aimed at what actually interests you.
Start simple, then tighten. First prompt without jargon. Second with a request for numbers or the aspect that interests you. You don’t need to write long instructions — the model handles simple questions well if the context is clear.
An LLM doesn’t replace sources. It helps you understand sources you’ll find yourself. And it helps you ask the right question before you start looking at all.
An article is a starting point. A conversation with a model is exploration. Where you end up depends on what question you bring.
This article is part of the Human–AI Collaboration category at AI907. Source articles: Regolith Shield: The 500-Ton Elephant in the Room and The Conventional Space of Human Life.
