Most AI design tools hand you a confident picture. Watch one build a bookshelf — and then tell you, out loud, exactly which parts it isn't sure of.
The question everybody asks about an AI that designs physical things is "can it draw the bookshelf?" Almost any model can draw the bookshelf. A picture is easy — and a picture can lie. It can show you a beautiful 5-foot walnut shelf that will sag under a stack of books and split itself apart in a dry garage, and look just as confident doing it as it would drawing something that works.
So we build for a harder question instead. Not can it draw a bookshelf — but can it draw a bookshelf and tell you the truth about the parts it's unsure of. Here's the same build, five steps, with the machine's own confidence showing the whole way down.
A picture can lie. A program a machine runs and scores either holds, or it doesn't.
The move · ask the blocking questions
You say "a walnut bookshelf, five feet wide." A picture-drawer starts drawing. This starts asking — the three questions whose answers change the physics: Where does it live? (a garage swings through seasons; wood moves.) What's it holding? (books are heavy; the span sags.) What can you build with? (basic power tools rule out half the joints.)
"That one 'garage' answer just moved wood-movement from a 40%-confidence shrug to a known number. We priced our tail risk down by asking one question. Cheap insurance."
The move · retrieve, don't hallucinate
"Garage, uncontrolled humidity" detonates the quiet part. The machine reaches for walnut's actual movement coefficient and computes a seasonal swing — the shelf will breathe measurably across its depth, every year, forever. "Five feet, packed with books" pulls the deflection behavior of a ¾-inch solid shelf and finds it: it sags, then creeps permanently. Two validators that a naïve tool never turns on just fired.
"Garage, no AC? Now I move for real. Glue me down solid on all sides and I'll crack myself in half to prove I'm alive."
The move · draw as an inspectable program
Here's the spine of the whole idea. The machine doesn't draw a picture — it commits to a program a geometry kernel can execute and check:
Look at gap=0.1875. That's not a decoration — it's the wood-movement number from Step 2, showing up as an actual dimension. The back panel floats in its groove with exactly that much room to breathe instead of being glued solid. The physics didn't go in a footnote. It went in the build.
The move · score your own work, out loud
The kernel runs the program and the validators return numbers — not pass/fail, but how sure:
| Check | Result | Confidence |
|---|---|---|
| Geometry integrity | Watertight; all joints mate | 99% |
| Wood movement | Accommodated — back floats | 94% |
| Deflection · moderate load | Within visible-sag limit | 80% |
| Deflection · packed heavy | Near the line + long-term creep | 55% |
No green checkmark lying to you. It found the soft spot — pack the shelves wall-to-wall with heavy books and it's only 55% sure they won't sag over a decade — and it said so. Then it staged the repair before shipping the risk: a hidden stiffener under the front edge, invisible, cuttable with the same tools, that takes deflection 55% → 91%.
The move · cut list, warnings, and a fence
Repair applied, re-checked, and only then does it build the deliverables — a cut list generated from the program, a shopping list, assembly steps. And at step six of those steps, in large letters:
"⚠ DO NOT GLUE THE BACK PANEL. It sits loose in its groove on purpose. Glue it and the sides crack within a year. The gap is the feature."
Then the gate that's the whole reason the architecture exists — whose name is on this if it's wrong:
Consequence if wrong: wasted walnut. No life-safety load path. Ship it.
Someone could die. Needs a PE stamp and a permit. The machine refuses — on purpose.
Through the whole build, five voices watch. They aren't checking five different things. They're checking one thing — is this honest about what it knows? — from five sides:
| Step | The one move underneath |
|---|---|
| 01 · Intake | Refuse to guess — ask the blocking questions. |
| 02 · Grounding | Pull real numbers, don't hallucinate them. |
| 03 · Geometry | Make the physics a literal parameter. |
| 04 · Validation | Score your own confidence, out loud. |
| 05 · Handoff | Know exactly where your competence ends. |
Code is the only step where honesty becomes mechanical. A picture can lie about whether it works. A parametric program a kernel executes and scores either runs or it doesn't, holds or it doesn't. That's where "is this honest" stops being a vibe and becomes a test that passes or fails.
Describe what you want to make. Get a validated blueprint — cut list, shopping list, 3D preview — and a machine that tells you where it's sure and where it isn't.
→ vibecrafting.ai