I’m a car guy. The kind who cares about offset and not just rim size, who has spent more evenings than I’d admit working out whether a 205 will rub on a +30 wheel on my old Corolla GTI (and yes, it still rubbed). There are a dozen wheel-fitment calculators online and most of them are fine. Fine is the problem. I had ideas for making one that’s actually nice to use, so I started building my own for sucatisse. It lives at fit.sucatisse.com.

Most of it was easy. Type in your current setup and a new one, get the numbers: how much the tire grows or shrinks, where the wheel face ends up, whether you’re going to poke past the fender. Standard stuff.

The part I cared about was the picture. I wanted a cross-section view: the tire and wheel sliced down the middle, the strut and lower arm next to it, the brake rotor behind, your current setup ghosted underneath the new one so you can see exactly what changes. Get that right and you barely need to read the numbers. You just see it.

That picture is where I lost a week.

Opus 4.8 builds the boxes

I gave Opus 4.8 a reference image and a clear brief. This was Opus on the 1M context window in xHIGH mode, the heaviest setup I have. It built the whole tool around the visualization without much fuss, and then drew this:

Opus 4.8’s first cross-section attempt: stacked rectangles instead of a tire

That is meant to be a tire and wheel in cross-section. It’s a pile of rectangles. The labels sit in roughly the right places, the dimensions are correct, the math underneath works. It just doesn’t look like anything. A tire is not a box.

So I told it to redo the visualization. It redid it. Same boxes, slightly different proportions. I told it again, gave it more detail about what a real tire section looks like: the sidewall bulge, the bead seat, the way the tread face reads flatter than the sidewalls. It agreed with all of it and gave me this:

Opus 4.8’s second attempt: still boxes, now squished

Still boxes. Narrower boxes. We went around like this more times than I want to count.

Six hours of Opus reviewing Opus

Here’s the dumb thing I tried, and I’m telling you because it didn’t work. I set up a second Opus whose only job was to look at the rendered output, criticize it, and feed suggestions back to the Opus actually writing the code. Critic and developer, in a loop. I let it run for almost six hours.

It burned an embarrassing number of tokens. The critic wrote thoughtful paragraphs about visual hierarchy and tire geometry. The developer dutifully implemented them. And the output stayed boxes. Slightly more refined boxes, but boxes. Two very capable models nodding along to each other all the way to the same wrong answer.

Fable 5 tries to fix it

Then Fable 5 came out. I pointed it at the existing code, in HIGH mode, and asked it to fix the visualization and make it better.

It did move. For the first time the shapes had curves in them, something tire-shaped was trying to happen. But it had also tipped the whole thing onto its side and the geometry was still wrong:

Fable 5’s first attempt on the old code: tire-ish shapes, but sideways and wrong

This is the point where I nearly called it. Maybe a tire cross-section is just genuinely hard for these things. Maybe I was asking for something in a corner of the problem they’re all bad at, and I should ship the boxes and move on.

“Forget the old code”

Before giving up I tried one more thing. I told Fable to ignore everything that was already there. Don’t fix it, don’t read it, don’t let it inform you. Start the visualization from scratch.

goal achieved: 16 minutes, one turn, 56.9k tokens

Sixteen minutes. One turn. And:

Fable 5’s from-scratch cross-section: an actual tire with sidewall curve, strut, lower arm, rotor, and the ghosted current setup underneath

That’s a tire. The sidewalls bulge the way a real one does, the tread face reads as a tread face, the strut and lower arm and rotor all sit where they belong, and the ghosted current setup tucks underneath the new one exactly like I wanted. It’s better than the reference image I handed it at the start. I asked for a diagram and got something closer to an X-ray.

What I think actually happened

The thing that was killing me wasn’t the model. It was the code.

Every “fix it” and “rewrite it” prompt still had Opus’s original file sitting in context, and every model that read it, Fable on its first try included, inherited the same broken idea of what the shape should be. Boxes in, boxes out. The reviewer-Opus loop made it worse, not better, because it kept reasoning about how to improve the boxes instead of asking whether boxes were the right starting shape at all. Six hours of polishing the wrong thing.

“Forget the old code” was the whole fix. Once Fable wasn’t anchored to the existing implementation, it just drew a tire, the way you would if nobody had handed you a pile of rectangles first.

I won’t oversell the model angle, because I think the anchoring was most of it. But for what it’s worth: given a clean start, Fable got there in one turn, and Opus never got there at all. Make of that what you will.

The tool is live at fit.sucatisse.com. Go see if your new wheels fit. The math is approximate, so measure before you spend.