Text to CAD: how AI turns
a sentence into a 3D model.
You type "a Japandi wardrobe wall, 3.2 m wide, three modules: hanging, shelves, drawers" and seconds later you have a parametric 3D model, a photorealistic render, and a STEP export. Here's what actually happens between those two moments.
$1.20 in free credits · No credit card required
What "text-to-CAD" actually means
Text-to-CAD is the workflow where a natural-language description becomes a parametric, dimensioned 3D solid model — the kind you can edit, measure, and export to STEP or DXF. It is not the same as text-to-image (which makes pixels) or text-to-3D (which usually makes unstructured polygon meshes).
The trick is that the AI does not draw your model. It calls geometry construction tools — add a box of these dimensions at this position, add a cylinder of this radius, stack panels in this configuration — and a separate engine executes those calls to build the actual model. The output is structured, editable, and exportable.
The pipeline, step by step
What happens between your prompt and your STEP file.
The model parses your intent
A large language model (in Prompt2CAD, Google Gemini 3) reads your prompt and a system specification that defines the design domain — units, coordinate system, available primitives, and what a valid model looks like. The model decides what the piece is, what its main components are, and what dimensions make sense for the category.
It calls geometry tools, not draws pixels
Instead of generating an image, the model emits structured tool calls: addBox, addCylinder, addSection, addSphere, with explicit positions, sizes, and materials. Each call is validated against a schema. This is where text-to-CAD diverges from text-to-image — the output is parametric geometry, not an array of pixels.
A builder turns calls into a real 3D model
A client-side builder API executes the tool calls against a Three.js scene. Coordinates are in meters, primitives are positioned at their center point, and materials map to physically based shaders. The result is a true 3D model you can rotate, measure, and export.
You refine through conversation
Because the model is parametric, follow-up prompts ("make it 10 cm wider", "add a drawer at the bottom", "swap to walnut") map back to specific tool calls and dimensions. The model edits the spec, the builder rebuilds, and you see the result in seconds. No undo stack, no manual modeling.
Geometry-grounded rendering
For photorealistic visuals, the geometry is sent as a depth or wireframe reference to an AI image generator like Nano Banana. The renderer is constrained by your actual model rather than guessing proportions from a text prompt. The picture matches what gets built.
Export to industry CAD formats
The final geometry is converted into STEP (B-rep solid for Fusion, SOLIDWORKS, FreeCAD), DXF (2D drawings for AutoCAD and CNC), OBJ and STL (meshes for visualization and 3D printing), and GLB (web 3D viewers). Real millimeters, real solids, ready for downstream tooling.
A worked example
From sentence to tool calls to solid model.
1. Your prompt
"A floating oak bookshelf, 1.6 m wide, 4 shelves evenly spaced, 30 cm deep, with vertical dividers between sections."
2. Tool calls the model emits
addSection({
id: "back-panel",
width: 1.6, height: 1.2, depth: 0.018,
position: [0, 0.6, 0],
material: "oak"
})
for (let i = 0; i < 4; i++) {
addSection({
id: `shelf-${i}`,
width: 1.6, height: 0.018, depth: 0.30,
position: [0, 0.24 + i * 0.30, 0.16],
material: "oak"
})
}
addSection({
id: "divider-left",
width: 0.018, height: 1.2, depth: 0.30,
position: [-0.4, 0.6, 0.16],
material: "oak"
})
// ... and so on3. What you see
A parametric 3D bookshelf appears in the viewport. Width, shelf count, and depth show up as adjustable parameters. You can rotate it, dimension it, render it with Nano Banana on a brick wall, and export STEP all from the same tab. Ask for "5 shelves instead of 4" and the model rebuilds in two seconds.
Why this matters in 2026
What text-to-CAD changes about the way furniture and built-ins get designed.
It collapses the early stack
Sketch, block-out in SketchUp, render in Nano Banana, export to STEP — four tools become one browser tab. The slow part of furniture design is the first 80% of the model, and that part is now seconds long.
It removes the blank canvas
Most design time is spent staring at an empty viewport deciding what to build. Text-to-CAD gets you to a tangible, editable starting point in one prompt, then lets you iterate from there.
It produces buildable geometry
Mesh-based AI 3D tools generate beautiful but unbuildable polygon soup. A parametric builder produces clean solids with real dimensions, ready for CNC cut files and manufacturing handoff.
It pairs with the tools you already own
STEP exports open natively in Shapr3D, Fusion 360, FreeCAD, and SOLIDWORKS. Use text-to-CAD as the concepting layer, then refine joinery and details in your usual editor.
What text-to-CAD cannot do (yet)
Honest limits, so you know where the edges are.
Rectilinear is solved, organic is not
Wardrobes, kitchens, bookshelves, desks, built-in storage — anything box-and-panel — works well today (Prompt2CAD's eval suite scores 0.88–0.99 on these categories). Curved legs, sculpted forms, and upholstered pieces still need a Rhino or Blender pass.
Spatial precision has bounds
AI is excellent at proportions, weaker at sub-millimeter joinery. Use text-to-CAD for the volumes and configurations, then tighten tolerances in a parametric editor before going to a CNC.
It is not a replacement for engineering CAD
FEA, motion studies, complex assemblies, and PLM workflows still belong in SOLIDWORKS or Fusion. Text-to-CAD is a concepting and visualization layer, not a manufacturing engineering suite.
Frequently asked questions
Is text-to-CAD the same as text-to-3D?
No. Text-to-3D tools like Tripo or Hunyuan generate polygon meshes from a prompt — visually impressive, but not editable, not dimensioned, and not useful for manufacturing. Text-to-CAD generates parametric solid geometry with real units, exportable to STEP and DXF for downstream CAD and CNC use.
What model powers Prompt2CAD?
Prompt2CAD uses Google Gemini 3 Flash for tool-calling and design reasoning, and Nano Banana for geometry-grounded photorealistic rendering. The model emits structured tool calls (addBox, addSection, etc.) that a client-side builder executes against Three.js.
Why not just describe your design to Nano Banana directly?
Image generators are trained on pixels, not geometry. Without a 3D reference they invent proportions, hardware, and structure that rarely match your spec. Prompt2CAD generates the model first and then uses it as a structural reference for the renderer, so the picture matches the design.
Can I edit the model after the AI generates it?
Yes, in two ways. Conversationally — "make the drawers deeper", "swap to oak" — and parametrically through dimension sliders for width, height, shelf count, and module configuration. For deeper edits, export STEP and refine in Shapr3D or Fusion.
What CAD formats can I export?
STEP (.step) for solid CAD interchange, DXF (.dxf) for 2D drawings and CNC, OBJ (.obj) and STL (.stl) for meshes and 3D printing, and GLB (.glb) for web 3D viewers. PNG screenshots of the viewport are also available.
How accurate are the dimensions?
All geometry is built in real-world meters. The AI is excellent at proportions and category-typical sizes. Sub-millimeter joinery tolerances should be verified or adjusted in a parametric CAD tool before manufacturing.
See it in action
Type a description, watch the parametric model build itself, render it, and export STEP. Free to try.
Start designing free →$1.20 in free credits · No credit card required