Text-to-CAD: parametric source, geometry evidence and responsible handoff
Text-to-CAD explained: source models, inspectable geometry and a separate manufacturing decision
Understand the agent-skill library and cadgen runtime, including why a rendered object is not proof of correct dimensions or print readiness.
What you will learn
- The output is an engineering artifact, not just an image
- Keep source, geometry and presentation distinct
- Define the evidence you need before generating
Before you start
- Basic Python functions, modules and file paths
- Understanding of units and the difference between a design and a physical part
Plan a bounded CAD exercise and distinguish source freshness, document inspection, visual review and manufacturing decisions.
Key takeaways
- CAD intent needs dimensions and relationships, not appearance alone.
- Source, document, preview and manufacturing readiness are different states.
- A selector-resolution result is not full geometry validation.
The output is an engineering artifact, not just an image
At revision 3e4dfde, earthtojake/text-to-cad provides agent skills for CAD, local viewing, drawings, robot descriptions and fabrication handoffs. The CAD workflow authors parametric Python models and uses cadgen to produce and inspect artifacts. The agent interprets a brief; the runtime performs geometry work.
A concept illustration can communicate an appearance without specifying exact dimensions. A CAD task needs explicit units, parameters, relationships and intended outputs. This series uses a harmless desktop sample block to explain that distinction; it does not design a load-bearing part, robot actuator or certified component.
Keep source, geometry and presentation distinct
A model script is the editable source when one exists. STEP is the default CAD output in the skill, while declared STL, 3MF and GLB outputs serve mesh-oriented uses. Mesh-only models are also supported. Imported STEP files can be inspected directly without pretending that a source script is available.
The viewer displays existing documents and their compiled representation; it does not run a Python model or determine whether the document is behind its source. A polished preview therefore cannot prove source freshness, dimensional correctness or suitability for a particular manufacturing process.
Define the evidence you need before generating
The CAD guide separates reference resolution, geometry soundness, targeted measurements and visual review. Its refs command can report successful selector resolution even for geometry that needs a separate validation check. A useful handoff states which exact file was checked and which requirements were actually measured.
We inspected seven fixed sources and exercised the upstream lazy-import proxy in eight isolated cases using a fake geometry module. No CAD kernel, model build, viewer, slicing, printer operation or performance benchmark ran. Those limits are preserved throughout the series.
Decision guide
| Criterion | Option A | Option B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
Implementation steps
- 1
Classify the requested artifact and intended use.
- 2
Record units, dimensions and acceptance checks.
- 3
Separate generated files from editable sources.
Copy-ready example
{
"example": true,
"part": "desktop sample block",
"units": "mm",
"geometryBuilt": false,
"dimensionsMeasured": false,
"manufacturingApproved": false
}Frequently asked questions
Is Text-to-CAD an image generator?
Its CAD workflow creates and inspects geometry artifacts; a render-only concept is a different task.
Does a good preview prove the part can be manufactured?
No. Geometry, process, material and fit evidence need separate checks.
Sources
- Text-to-CAD / README.mdSource checked 2026-09-14
- Text-to-CAD / skills/cad/SKILL.mdSource checked 2026-09-14
- Text-to-CAD / skills/cad-viewer/SKILL.mdSource checked 2026-09-14