Text to CAD Skill Family
An exact-revision, locally installed family for CAD, robot descriptions, slicing, fabrication preflight, and cautious printer handoff. It fills a previously absent workflow domain rather than duplicating an existing skill. Source:
earthtojake/text-to-cadcommit2a3a42a39fb9667e551386b55c2120774e35497e, 2026-08-12
Installed routes
| Job | Executable skill |
|---|---|
| STEP-first parametric parts and assemblies | cad |
| Visual inspection across CAD and robot formats | cad-viewer |
| Canonical off-the-shelf part discovery | step-parts |
| Two-dimensional manufacturing drawings | dxf |
| Robot structure and kinematics | urdf |
| MoveIt planning semantics | srdf |
| SDFormat models and worlds | sdf |
| Experimental browser-native implicit geometry | implicit-cad |
| Printer-profiled slicing and static validation | gcode |
| SendCutSend upload/order preflight | sendcutsend |
| Bambu LAN upload and explicit print start | bambu-labs |
The canonical copies live under skills/engineering/ with upstream license and support files. The versioned CAD and Fabrication Workflow selects the relevant subset, requires a brief/units/assumption ledger, preserves source artifacts, and separates geometry generation from physical execution.
Safety and proof
STEP is the primary editable artifact for ordinary CAD; mesh exports are secondary. Implicit CAD is experimental and never silently replaces the STEP-first route. Robot descriptions require frame, joint, limit, inertial, collision, and planning-semantics validation rather than plausible XML. Slicing requires an explicit printer/profile wrapper. SendCutSend remains a preflight report. Bambu traffic defaults to dry-run; uploading requires --execute, and starting a print additionally requires --confirm-start-print plus user authority. Source: pinned skill manifests
At capture, the MIT repository had 13,286 stars, 1,401 forks, 1,870 blobs, 12 skill manifests, 106 test files, and self-contained packages for CAD generation and viewing. Those tests support the source's engineering maturity, but this machine still needs per-skill dependency doctors before claiming every path is runnable. Source: frozen repository evidence, 2026-08-12
Model, render, compare
Wes Bos modeled a whole house in Three.js by capturing iPhone LiDAR with hundreds of photos, then running a skill that models each object and checks the model against the photo taken from the same spot and angle, repeating until they match. The check is what makes the result dimensionally accurate. Skill, viewer and scanner app are in wesbos/3d-model-anything. Source: X/@wesbos thread, 2026-10-22; Source: YouTube, "my office is complete"
nurb applies the same loop to printable parts: the user describes a part, the agent writes CAD, and a live viewer shows it before printing. Source: nurb.dev; Source: Shpigford/nurb
The rule for cad work: render from the reference viewpoint and compare against the reference, every pass. A model the agent never rendered is a guess. The general prompt is Compare Against the Reference.
Timeline
- 2026-08-12 | Read every skill manifest, the package/viewer docs, representative tests, release workflow, and license; installed the 11 relevant canonical skills at the pinned revision; added deterministic routing and physical-action gates. Source: X bookmark
2084114545690423407; pinned repository evidence