The Part That Doesn't Exist

I'm in the middle of a whole-home lighting rewire. Lutron RadioRA 3 throughout. New keypads, dimmers, switches — the works. And in one hallway, I hit a problem that would have derailed the project for weeks a few years ago.

The wall where the keypad needs to go isn't flat. Or rather, it's flat, but it meets another surface at an angle. The electrical box is recessed normally, but the finished face needs to sit at a roughly 24-degree tilt so the keypad reads correctly from the hallway. The shallow side of the box extension needs to be about 1-1/8 inches deep. The deep side, 3 inches.

This part does not exist. No manufacturer makes an angled 2-gang electrical box extension at this depth ratio. You can't order it. You can't find it at Home Depot. There's no SKU for "the specific geometry of your specific wall."

Before 3D printing, you'd have three options: pay an electrician to fabricate something custom out of sheet metal, build a janky wooden shim and hope it passes inspection, or relocate the box entirely and patch the drywall. All of them cost more time and money than the part itself warrants. So historically, you'd compromise. Move the switch somewhere less ideal. Accept the angle. Live with the thing that's slightly wrong in your house forever.

I didn't do any of that. I sat down at my desk, opened a conversation with Claude, described what I needed, and had a printable STL file within an hour.

[IMAGE: The angled hallway wall, showing where the electrical box sits and why the geometry is unusual. A wide enough shot to show the spatial problem — two surfaces meeting at an angle that no standard part accommodates. Before any fabrication. The reader needs to see the problem before the solution.]

What actually happened

I started with a rough SketchUp model I'd been noodling with, but I couldn't get the geometry right. The compound angle between the angled face and the device mounting tabs was tripping me up. So I described the problem in plain language: a 2-gang box extension, open back, angled front face with mounting ears for a standard cover plate, two 1-gang switch positions with tabs perpendicular to the angled surface.

The first thing Claude did was ask clarifying questions. Not vague ones — specific ones that would change the model. Is this flush-mount or surface-mount? Does the cover plate sit flush to the angled face, or is there a separate flat mounting surface? These are the questions an experienced electrician or industrial designer would ask, and they're the ones a beginner doesn't know to answer until someone forces the issue.

Then it wrote the entire part in OpenSCAD — a parametric modeling language that exports directly to STL for 3D printing. Every dimension was a named variable at the top of the file. Wall thickness. Flange depth. Screw spacing. Gang center-to-center distance. If I needed to change anything, I'd change one number and re-render.

[IMAGE: Screenshot of the OpenSCAD render or the parametric code with named variables visible. If both exist, prefer a split or side-by-side showing code on one side and the 3D preview on the other. The reader should see that the part is defined by editable parameters, not fixed geometry. This is the "build guide" moment — it makes the method legible.]

The first version was close but wrong. The interior clearance on the shallow side was too tight. When you tilt two standard receptacle bodies 24 degrees inside a box, they shift laterally toward the deeper wall. The bodies would have collided with each other and with the box walls. This is not a problem you can see in your head. You need the math.

Claude caught it — or rather, I asked if the receptacles would actually fit, and Claude ran the geometry. It modeled the physical envelope of a standard receptacle body (31.75mm wide) rotated at the tilt angle, checked the lateral displacement, and found 2.4mm of clearance on the right side. Technically it fits. But "technically fits" and "a licensed electrician would feel comfortable with this" are different things.

Three options came back: reduce the tilt angle (solves clearance but defeats the purpose), widen the box beyond the standard footprint (works but the back plate overhangs the existing box), or accept the tight clearance. I chose to widen the box. Claude adjusted the parametric model, re-checked clearances, regenerated the STL.

Then I asked to change the depths — 35mm and 83mm instead of the original spec. One parameter change. Rebuilt in seconds. New STL, new renders, same verified clearances.

I printed it on a Bambu Lab X1 Carbon in PETG. It took about three hours. It fit perfectly on the first try.

What the AI actually did

It's worth being precise about what happened here, because the discourse around AI and making tends to oscillate between "AI will replace all craftspeople" and "AI is useless for anything physical." Both are wrong, and both miss the interesting part.

Here is what the AI did not do: it did not hold a drill. It did not run wire. It did not measure the wall. It did not decide where the keypad should go. It did not know that this hallway is the one my daughter walks through every morning, and the light switch needs to be reachable and visible from a specific sightline. Those are human decisions grounded in lived experience with the space.

Here is what the AI did: it translated my intent into precise geometry. It knew the NEMA standards for 2-gang box dimensions, device screw spacing, and receptacle body envelopes. It wrote parametric code that I can modify forever. It ran clearance checks I would have had to do by hand with calipers and trigonometry. And it did all of this in a conversational loop where I could push back, change my mind, and refine the design without starting over.

The closest analogy is working with a very knowledgeable colleague who happens to be fluent in OpenSCAD and never gets impatient when you change the spec for the third time. The difference is that this colleague is available at 11pm on a Tuesday when you're standing in a half-finished hallway with a headlamp on, trying to figure out if your idea is even physically possible.

[IMAGE: The finished installation — keypad mounted, cover plate on, the angled surface reading cleanly in the hallway. The custom part is invisible. That is the point. Shoot it from the sightline the daughter would see walking down the hallway. The reader should understand that the geometry serves a specific human path through the house.]

The gap this closes

The barrier to making physical things has never been the tools. You can buy a 3D printer for $200 that would have been a $50,000 machine fifteen years ago. You can watch a YouTube video on how to wire a 3-way switch. The raw capability is there.

The barrier is the space between intent and specification. You know what you want — "I need this switch to sit at an angle" — but you don't know how to describe it precisely enough to fabricate. You don't know the standards. You don't know the clearances. You don't know what you don't know, which is the part that kills most projects before they start.

AI closes that gap. Not by replacing your judgment or your hands, but by making the translation from "I want this" to "here are the exact dimensions, and here's why they work" dramatically faster and cheaper. The part still has to be printed. It still has to be installed. You still have to live with it and find out whether it actually works in practice. But the design phase — the part that used to require either deep expertise or expensive professional help — is now a conversation.

My daughter will never know that the light switch in her hallway required custom geometry. She'll just reach for it and the lights will come on. That's the point. The best-made things in a home are invisible. They just work, exactly where they should be, shaped to fit the life that happens around them.

The part that doesn't exist is the one you make.