Process / the laptop stand
Two designers,
one brief,
three rounds
The laptop stand was not designed once. It was designed twice in parallel, by two different AI models that never spoke to each other, then judged at a bench by a person who had to live with the result.
This page is the account of how that went, including the parts that went badly — because the failures are where the method earned its keep.
The bench
Sets the brief, judges every round, and is the only one who can say a thing is good. Has to actually use the part.
Claude
Writes the brief, runs both designers, and independently re-measures everything they claim before it reaches the bench.
Grok
One of the two competitors. Given the brief and the existing part; no sight of the other's work until the judging.
Codex
The other competitor. Same brief, same files, same deadline, no coordination.
The rule that makes it work
Nothing either designer says about its own work is taken at face value. Every part is re-rendered from source, its bounding box and solid volume re-measured off the exported mesh, and its claims re-probed independently. A design that reports a clearance has to survive somebody else intersecting a solid cone with it. That is the whole method: disagreement to generate options, measurement to settle them, and a human at every gate.
Round one · make the compartments functional
They both refused the brief, in the same words
The existing part had ten storage openings and 201 ml of volume, and the complaint was that they were generic rectangular holes. Both designers were told to make every compartment actually be for something. Both came back having rejected the premise that more volume was the goal — independently, and almost identically.
“201 ml of miscellaneous holes is not a target worth retaining.”Codex
“That 201 ml was leftover plastic with a number on it.”Grok
Where they split was depth. Codex said a deep pocket is a trap and rebuilt everything as shallow trays with sweep ramps, dropping the part's weight by 40%. Grok said depth is how a thing stands proud so you can grab it — “a Sharpie in a scooped tray is worse than a Sharpie in a bore” — and split the block's height instead: tall where something stands upright, lower where a hand has to reach inside.
Grok's split-height idea won and became the architecture. Codex's cable rail won separately: a rounded throat with a full-width opening to the top, so a lead drops in and lifts out, against Grok's inherited design where the cable had to spring past a narrow neck. Two reasons it was the better answer: no thin spring fingers to fatigue in PETG, and it still works with the bench against a wall.
The fillets had never existed
Reviewing the two entries turned up a defect in the shipping part — one that had been live on this site. The fillets at the base of the uprights, the only load-bearing joint in the whole thing, were placed one quadrant wrong in the source and had all landed buried inside the base slab, doing nothing. The part had square roots and the drawing said otherwise.
region a root fillet should occupy Simple: no Vertices: 6 <- a bare coplanar face. no solid. nothing there.
Worse, the original clearance test had passed — and passed because the fillets were missing. A correctly-placed fillet would have fouled the laptop and the test would have caught it. A green result for the wrong reason is more dangerous than a red one.
Both designers were made to fix it, and both then found something the brief had got wrong: an interior fillet eats into the space the laptop occupies. Each independently kept the inside blend smaller than the felt liner so the machine still bears on the slot floor. Both also worked out that the probe being used to check them was itself unreliable — it shared faces with the part, which makes the geometry engine report a false negative — and re-ran it inset. They were more careful than the instruction was.
Round two · two different parts
An objection became the specification
In its first write-up, Grok had pushed back on a constraint: the brief forced both cradles to be the same part, and it thought that was wrong.
“Two identical cradles doubles everything. The pair holds four SD cards, eight microSD, two headers, four USB sticks. That is more card storage than this bench will fill. Two different parts would be the real answer; the brief forbids it.”Grok, round one
The bench agreed and lifted the constraint. That is the part worth noticing: the losing argument in one round became the requirement in the next. An AI designer that only ever complies cannot do that for you — you need one that files an objection, and a process that reads it.
Round three · the fact that settled it
Every port is on one side
Up to this point the two cradles differed by taste — one got fasteners, the other got cards. Then a look at the actual machine settled it. Every port on this laptop is on one side. Standing on edge, that means every cable leaves from one end of the pair.
So the two parts are not a matter of preference. One is the cable cradle: it sits at the port end and carries the whole machine's cabling in seven open lanes, each sized to its own lead. The other is the dry cradle: no cable will ever reach it, so the entire back foot it would have spent on a rail goes to storage instead.
That one observation made Grok's pair lighter and roomier at the same time — storage up from 163 to 195 ml while the plastic went down from 597 to 566 cm³ — purely by deleting a rail that would never have seen a cable.
| Where it ended up | Grok | Codex |
|---|---|---|
| Storage compartments | 22 | 17 |
| Storage volume, the pair | 195 ml | 155 ml |
| Solid volume, the pair | 566 cm³ | 497 cm³ |
| Cable lanes | 7 | 7 |
| Driver bore sizes | two | one |
The last change came from the bench, not from either model. Codex had put a second cable comb immediately behind the laptop. It was rejected on sight: the cables come out of the machine's end, never its back, so no lead would ever route there — and with the laptop in place you cannot get a hand into that strip anyway. It became two more storage bins.
What the competition actually bought
Not “better output.” Four specific things, each traceable to a decision:
Independent agreement is evidence
When two models that cannot see each other's work reject the same premise in the same terms, that is a much stronger signal than either one saying it alone.
Objections carry forward
The requirement that defined the final part — two different cradles — started life as a designer's complaint about the brief in a round it did not fully win.
They cross-pollinate
The shipped pair is Grok's architecture with Codex's cable rail. Codex's own second entry abandoned its footprint and adopted its rival's, in writing.
Verification beats trust
Every dimension here was re-measured off the exported mesh rather than read from a write-up, and the browser model's 66 layout figures are diffed against the real source automatically. The defect that mattered most was found by measuring, not by looking.
Failure is cheap, so ask for more of it
A round costs a brief and some compute. That is what makes it reasonable to throw away a whole design direction because a shape of storage was wrong.
A person holds every gate
Nothing here was adopted because a model recommended it. Each round ended with someone who has to use the part saying yes, no, or that makes no sense.