One phone photo and two ruler measurements were enough to get a printable replacement for the winch hook our twins lost off their RC Jeep. No 3D scanner, no 360° photo session, no CAD skills.
The lost hook
We bought the twins matching remote control Jeeps, and each one came with a little black winch hook on a cord. Within a few weeks one of the hooks was gone. Nobody knows where, and the carpet isn't telling.
The surviving hook is a simple part: an eye at the top for the cord, a thick shank, and a J-shaped curve at the bottom. It's about 2.5 cm tall and 1.5 cm wide at the widest point. It looked like exactly the kind of thing a 3D printer is good at, if I could just get a model of it.
Why I skipped the 3D scanning apps
The scanning apps I knew about use photogrammetry. You walk around the object taking dozens of photos from every angle, and the app stitches them into a 3D mesh. That works well for complex, organic shapes. For a 2.5 cm piece of shiny black plastic, it's a tough job: small, dark, glossy objects are some of the hardest things for those apps to capture cleanly.
More importantly, it was overkill. This hook is essentially a flat profile with rounded edges. Everything that makes it a hook (the eye, the shank, the curve) is visible in a single photo taken straight on. The third dimension is just thickness. If I could trace the outline and give it a thickness, I'd have the part.
The approach: one photo, two measurements
I set the surviving hook on the carpet, took one photo with my phone from directly above, and put it in a folder on my computer. Then I asked Claude, Anthropic's AI assistant, for help:
The one and only input photo: the surviving hook, shot straight down.
Can you make a stl 3d printable model of the hook that I can replace a missing one from? This hook is about 2.5cm long at the longest and 1.5cm wide at the widest. It doesn't need to be pixel perfect... just roughly the same.
That was the whole input: the photo, the overall height and width, and permission to be approximate. Claude had access to the folder through the Claude desktop app, so it could look at the photo directly and write the finished files back into the same folder.
A few minutes later I had an STL file sitting next to the photo, plus a rendered preview to compare against the real thing.
Behind the scenes: what Claude actually did
The part I found most interesting was how the model got built. There was no tracing tool and no CAD program. The hook was described entirely in Python code, then converted to a mesh and checked like any other piece of software.
- Looked at the folder. It held the photo (
PXL_20261101_172606934.jpg) and a first version of a generator script,generate_hook.py, that was already in the folder. - Read the photo and the script. Claude viewed the photo and compared it to the shape described in the script, checking that the eye, shank and hook curve lined up with what the photo showed.
- Described the outline as math. The script defines the hook's 2D outline as a signed distance field: a function that tells you, for any point, how far it is from the edge of the hook (negative inside, positive outside). The outline is built from three simple pieces:
- a circle for the eye, about 8.6 mm across, with a 4.3 mm hole for the cord
- a smooth curved "spine" through 9 control points, each with its own thickness, so the shank swells to its widest at the belly and tapers toward the tip
- a circle subtracted from the inside of the J, so the throat of the hook stays cleanly open
- Gave it thickness. The flat outline was extruded to 4.5 mm thick, with 1.55 mm rounded edges all around. Both faces stay flat, so it lies on the print bed with no supports.
- Rendered a preview and compared it to the photo. The script draws a shaded front view on a 5 mm grid and a tilted 3D view. The silhouette matched the real hook closely enough that no shape changes were needed.
- Built the first mesh and checked it. The script's built-in mesher produced about 120,000 triangles. Claude then counted how many triangles share each edge. In a clean, closed solid, every edge is shared by exactly two triangles. This mesh had 12 edges that weren't, a small defect most slicers would quietly repair, but not something worth shipping.
- Tried a quick fix, then a better method. Nudging the sampling grid only reduced the bad edges to 4, then 1. So Claude wrote a short new script,
build_stl.py, that rebuilds the mesh with the marching cubes algorithm from the scikit-image library, then nudges every vertex exactly onto the hook's surface. - Verified the final file. The rebuilt mesh came out with 0 bad edges and consistent triangle orientation, which means it's watertight and slices cleanly. Final numbers: 15.0 × 25.0 × 4.5 mm, 290,840 triangles, about 790 mm³ of plastic.
- Saved the results back to my folder:
rc_jeep_hook.stl, the preview image, andbuild_stl.py.
Here's the core of the shape definition, the list of points the hook's spine passes through. Each row is an x and y position in millimeters plus the thickness at that point:
spine_ctrl = [
[5.65, 19.20, 3.30], # Inside eyelet base
[5.20, 15.20, 2.85], # Neck
[3.60, 10.00, 3.45], # Left belly (thickest part of the shank)
[4.10, 5.60, 3.30], # Lower left bend
[6.60, 3.15, 3.10], # Bottom-most bend
[10.00, 3.25, 2.85], # Lower right bend
[12.55, 5.45, 2.35], # Right curve rising toward tip
[13.45, 8.35, 1.80], # Upper right tip curve
[13.15, 10.10, 1.50], # Rounded hook tip
]And the check that caught the defect, simplified:
# Every edge of a closed solid should belong to exactly 2 triangles
edges = np.sort(np.concatenate([f[:, [0, 1]], f[:, [1, 2]], f[:, [2, 0]]]), axis=1)
_, counts = np.unique(edges, axis=0, return_counts=True)
print("bad edges:", (counts != 2).sum()) # first mesh: 12, final mesh: 0The nice thing about a shape defined this way is that it's easy to tweak. Want the hook 10% bigger, or a larger hole for a thicker cord? Change a number and rerun the script.
The result and printing tips
The finished model is 25 mm tall, 15 mm wide and 4.5 mm thick, with the same eye, shank and J-curve as the original. It isn't a pixel-perfect copy, but it doesn't need to be: it just has to hold the cord and look right next to its twin.
The generated model: front view on a 5 mm grid (left) and a tilted 3D view (right).
Setting | Recommendation |
|---|---|
Orientation | Lay it flat. Both faces are flat, so no supports are needed. |
Material | PETG or ABS if the winch pulls hard; PLA is fine for light play. |
Strength | 100% infill, or at least 4 walls. It's a small part, so this costs almost nothing. |
Size | Scale it in your slicer if your hook is a bit bigger or smaller. |
If your kids' Jeep uses a different hook, the same trick works for any part that's basically flat: one straight-on photo, the overall dimensions, and a request for an STL.




