I asked an LLM to help me figure out the stair design for my round house. I’ve got a 24-foot diameter, 10'8" wall height, center ring at 15 feet, and I need code-compliant stairs to a loft. California says max 7¾" riser, min 10" tread, 36" wide. The LLM gave me a clean L-shaped layout with storage under each run. Dimensions worked. Riser count was right. It even suggested lift-up treads for blanket storage.
I could have just started cutting.
Instead I sat down with the actual floor plan, reviewed the work, and realized the stair placement would block the kitchen flow, put the loft access in the wrong spot, and eat six feet of living space I couldn’t afford to lose in 452 square feet. The LLM answered the question I asked. It didn’t know the questions I hadn’t asked yet.
This is vibe coding, but for buildings. The stakes are just more obvious when the material is wood.
What Vibe Coding Actually Is
Andrej Karpathy coined the term earlier this year. The idea is simple: you describe what you want to an AI, it generates code, you run it. If it works, you move on. If it doesn’t, you describe the problem and it tries again. You don’t need to understand what the code does. You just need to describe what you want it to do.
For throwaway scripts and prototypes, this is fine. For anything that matters, it’s a trap. Not because the code is always wrong. Sometimes it’s remarkably right. The trap is that you lose the ability to evaluate it. You can’t tell good code from bad code if you don’t understand what you’re looking at. The feedback loop breaks.
The Same Trap, With Lumber
I’ve been building things on my property for the last year and a half. Welding, framing, wiring, running a chainsaw. I broke my wrist doing it. I also learned more about how structures actually work than I did in years of hiring contractors.
I’m planning a circular deck, 24-foot diameter, composite decking. To cut the circle, the right answer is a trammel, a rigid arm set to your radius that scribes the arc from a center point. You run your boards past the line, scribe them in place, cut them with a jigsaw. Every cut end needs to land on blocking underneath, which means you need a curved header framed to the same arc.
An AI could tell me the trammel method. It might mention the blocking. What it won’t tell me is that composite decking moves differently than wood in the sun, that the Camo hidden fasteners I’m using need at least 1.5 inches of bearing on each board end, or that a string trammel stretches enough on a 12-foot radius to produce a visible wobble. I’m learning those things from reading the manufacturer specs, from studying other builds, from talking to people who’ve done it.
The knowing matters. Not because the AI is wrong, but because when something goes sideways at 2pm on a Saturday with the saw running, you need to understand why the plan was the plan. You need to be able to adapt in real time. If all you have is the output and none of the reasoning, you’re stuck.
The Point Isn’t That AI Is Bad
I use AI tools constantly. I run an AI company. I used Claude to help work through the stair dimensions, the deck layout, the electrical planning for my house. When I was figuring out the deck boards, I needed a cut list that accounted for standard and grooved composite boards to maintain a picture frame border around a 24-foot circle. That’s tedious math with a lot of waste optimization. The AI knocked it out in minutes. I could have done it by hand. I’m glad I didn’t.
I was the one who knew what to check. I was the one who caught the kitchen flow problem. I was the one who knew which questions to ask next. The AI made me faster. It didn’t make me competent. I had to already be competent for the AI to be useful.
You can vibe your way to a prototype. You can’t vibe your way to a house that passes inspection and doesn’t burn down. You can’t vibe your way to software that handles edge cases and doesn’t leak data. At some point you have to actually know what you’re building and why.
Pick up the saw. Read the code. Move the dead tree yourself.