Your Chatbot Can Now Reverse-Engineer a Broken Part. Bring Calipers.
A clip snaps, a bracket cracks, a knob goes missing, and the replacement is either discontinued or two weeks out. That used to mean an evening in CAD, if you knew CAD. Now a growing number of people are pointing a general-purpose chatbot at the problem: a few photos, a few caliper readings, maybe a phone scan, and out comes a printable STL or an editable model.
It works more often than you'd expect. It also fails in very predictable ways, and almost all of them come down to one thing: measurements.
What people are doing
The most-shared version of the idea this fall is Greg Isenberg's "vibe manufacturing playbook" on X. Step one is not "ask the AI." It's "measure everything": digital calipers for small parts, a phone LiDAR app like Polycam for bigger objects. In his words, "Bad measurements are the #1 reason first builds fail." Step two is having a frontier model write the design as code.
Other examples point the same way. An AI News Reporter post relays a project by @toyoshi, where an iPhone scan of a broken concrete corner became a 3D-printed patch. GPT-6 Astra reportedly generated the print data without being told which section was missing, and the creator had no point-cloud experience. Industrial designer Robert Bye says a part that would have taken him a few hours in SOLIDWORKS took Claude minutes through the Fusion 360 MCP connection.
Not everyone is impressed. Dylan Zemek, working with GPT models and Fusion 360 MCP, wrote that the outputs "really aren't good, they overlap things, don't understand joins." And on Reddit's r/3Dprinting, a user got ChatGPT to output an STL from a photo that came out as a plain block, missing the hole. All of these are self-reported posts, so take them as signals, not benchmarks.
How it works, plainly
There are two routes, and they work differently.
The photo route. The chatbot isn't scanning anything. As Meshra explains, the photo is a hint. A vision model guesses the shape, then writes CAD code (CadQuery, OpenSCAD, or a Python script) and exports an STL. Meshra says outright that this "will not reproduce an exact replica." The real size comes only from what you give it: a ruler or coin in the frame, plus the numbers you typed in from your calipers.
The scan route. A phone or structured-light scanner gives you a triangle mesh, which is lumpy, full of holes, and not editable in any useful way. AI tools repair the mesh, then fit real features to it (flat faces, holes, extrusions) so you get a solid you can change, such as a STEP file or a Fusion feature tree.
Either way, you iterate. Print it, hold it next to the original, then tell the AI what's off, with marked-up photos and new measurements, and run it again. Bruno's Workshop sums the loop up as: measure, describe, generate, test-fit, repeat.
The tools
- Solid Vibes' broken-part playbook is the most practical how-to: shoot straight down with a coin or ruler for scale, give caliper values for the critical features, test-print just the mating interface first, and use PETG or ASA instead of PLA.
- OpenSCAD skill for Claude Code is open source. It generates OpenSCAD models, rebuilds STLs as SCAD code, replicates parts from photos, and checks printability. Because the output is code, you can edit a dimension and regenerate.
- fusion-reconstruct-mcp is an open-source add-in that rebuilds an STL or 3MF as parametric Fusion 360 features and compares the result with the source. Its own README calls the photo path "assisted modeling," not scan-to-CAD.
- Backflip does scan-to-parametric-CAD with a Fusion add-in. It announced on X that it also takes drawings and phone photos. Backflip lists "$10 a part," four free conversions, and plans from $20 a month. Those are the vendor's numbers.
- OriginShape has an AI copilot, Oscar, that turns scans into editable STEP or IGES files and flags features it's unsure of. Per its site, it launched Oct. 26.
- Zoo's Zookeeper is a chat-driven CAD agent. Reverse engineering from STEP or STL is described as still in training, not shipped.
The catches
A photo has no true dimensions. If the part has to fit something, the numbers have to come from calipers. No exceptions.
First tries miss things. Holes, chamfers, snap features, and tapers are the usual casualties, as the Reddit block shows. A scan can also faithfully copy damage, such as wear, warping, or a missing chunk, unless you point it out.
Blocky parts work best. Brackets, clips, spacers, and knobs are good candidates. Organic, freeform shapes are not.
Material matters. A PLA print of a part that lives in a hot car or takes a load will fail. Pick the material for the job.
Not for safety-critical parts. Brakes, lifting gear, pressure parts, and anything a person depends on are off the table.
Mind the IP. Copying a part for your own repair is one thing. Copying a patented or branded OEM part for resale is another.
Everything here is new. Some tools launched days ago, some features aren't shipped, and prices are whatever the vendor says this month.
Try it Monday
- Pick an easy one. Choose a simple, blocky broken part: a clip, bracket, or spacer, nothing curvy and nothing that holds anything up.
- Photograph it right. Shoot straight down on a plain background with a ruler or coin in the frame. Add a side view.
- Measure 2–3 critical dimensions. Use calipers on whatever mates with something else: hole diameters, slot widths, overall thickness.
- Ask for code, not just an STL. Ask ChatGPT or Claude for OpenSCAD or CadQuery with your measurements as named variables, so you can change one number later.
- Print only the fitting feature first. A thin slice with the hole or clip is faster and cheaper than the whole part.
- Iterate with marked-up photos. Circle what's wrong, give the corrected measurement, and regenerate.
- If you have a scan, try the scan route. Run it through fusion-reconstruct-mcp or Backflip's free conversions to get an editable feature tree, then check the critical dimensions against your calipers anyway.
The chatbot does the modeling. You still have to do the measuring.