Prompt & Image Guide

The quality of your prompt (for Text-to-CAD) or source image (for Image-to-CAD) is the single biggest factor in output quality. This page covers what the pipeline handles well, what language patterns produce better results, and what to avoid.

Text-to-CAD Prompt Guide

Geometry That Works Well

Makistry's pipeline generates parametric CAD using CadQuery, a Python-based B-rep modeling kernel. It excels at:

• Prismatic shapes: boxes, cylinders, extrusions, pockets, shells
• Standard features: chamfers, fillets, drafts, thin-wall shells
• Hole patterns: through-holes, blind holes, counterbores, countersinks, slots
• Threaded features: tapped holes — specify M-size (e.g. M5, M3 × 0.5)
• Assemblies of simple mating parts: enclosures, brackets, clamps, fixtures
• Standard hardware: fasteners, bearings, gears — Makistry has a parts warehouse; name them by size (e.g. "M6 hex bolt, 20mm length", "608 ball bearing", "M3 heat-set insert")

Language Patterns That Produce Better Results

Be specific with dimensions — always include units:

✓  A 50×30×10mm rectangular plate
✗  A small flat plate

Name the feature type explicitly:

✓  An M5 counterbore: 8.5mm head diameter, 4.5mm deep, centered on the top face
✗  A screw hole on top

Specify the reference face or location:

✓  Two M4 through-holes on the long face, 15mm from each end, centered vertically
✗  Two holes on the side

Describe mating intent for assemblies:

✓  The lid sits on a 1.5mm step that runs around the inner perimeter of the base, with 0.15mm clearance
✗  The lid goes on the base

Use geometric vocabulary:

✓  Circular pocket, 20mm diameter, 5mm deep, centered on the top face
✗  A round dent in the top

Three Example Prompts at Different Levels of Complexity

Hobbyist — Lens Cap

A lens cap for a 52mm filter thread: 60mm outer diameter, 52mm inner diameter at the 
retention lip (3mm deep), 3mm wall thickness throughout. Snap-fit retention ring on the 
inner circumference, 1mm wide × 2mm tall, continuous around the full inner circumference.

Engineering — Mounting Bracket

An L-shaped aluminum mounting bracket. Horizontal leg: 80mm long × 40mm wide × 4mm thick. 
Vertical leg: 60mm tall × 40mm wide × 4mm thick. Both legs share a 40mm width. Two M5 
through-holes on the horizontal leg: 10mm from each end, centered on the 40mm width. 
Triangular gusset on the inner corner: 20mm on each leg face, 4mm thick. All outer corners 
chamfered 1mm.

Assembly — Electronics Enclosure

A two-part electronics enclosure. Base: 100×60×35mm external dimensions, 2mm wall thickness, 
open top. Four M3 boss inserts at the interior corners: 6mm OD, 3mm ID, 5mm tall, positioned 
3mm from each corner. A 2mm stepped lip runs around the full inner perimeter at the top, 
1.5mm tall, to accept the lid. Lid: 100×60×12mm external, 2mm walls, closed top. Inner step 
around the perimeter: 2mm tall × 2mm deep, sized to seat on the base lip with 0.2mm clearance 
on all sides. Four M3 through-holes in the lid corners, aligned with the base inserts.

What's Not Supported

• Organic freeform surfaces: cloth, skin, terrain, splash forms, sculptural shapes
• Fluid or simulation geometry: anything that requires computational simulation to define shape
• Sub-0.1mm features: the kernel can represent them but the pipeline may not reliably produce them
• Assemblies with more than 8–10 distinct parts in standard or pro mode — switch to "assembly" mode, which parallelizes generation across parts
• Injection molding draft analysis, wall thickness analysis, FEA — geometry only, no simulation

Image-to-CAD Source Image Guide

What Makes a Good Source Image

The image pipeline extracts edges, interprets geometry, and reconstructs a 3D B-rep model. The quality of that reconstruction depends almost entirely on how clearly the geometry is visible.

Best sources, in order of accuracy:

1. Dimensioned engineering drawings — dimensions are read directly; highest accuracy
2. Hand sketches with labeled dimensions — nearly as accurate; much faster to produce than a photo
3. Orthographic photos (top, front, side) — good accuracy; supply multiple angles for 3D reconstruction
4. Perspective product photos — moderate accuracy; geometry is estimated from visible edges
5. Renders — similar to photos; useful when the physical part doesn't exist yet

Ideal Image Conditions

• Single part per image (or a clearly dominant part)
• Clean background — white, light grey, or contrasting solid color
• High contrast between part edges and background
• Even lighting — avoid shadows that obscure edges or create false edges
• Sharp focus throughout the image — especially at edges
• Orthographic projection preferred over perspective; angled photos require the model to estimate foreshortening

Using Multiple Images

Submit up to 3 images of the same part from different viewpoints. Pair with a prompt hint to identify each view:

image 1: part_front.png
image 2: part_side.png  
image 3: part_top.png
prompt: "front view, right-side view, and top view of the same bracket"

Multi-view submissions significantly improve 3D reconstruction accuracy, particularly for parts with features that are only visible from certain angles.

What Does NOT Work

Resolution is rarely the problem. A sharp 800×600 image of a clean part will outperform a blurry 4K photo. Clarity and contrast matter more than megapixels.

Image vs. Drawing: Behavioral Differences

If you have a dimensioned drawing, always use it over a photo for the same part.