From a Prompt to an STL: The AI Skill That Designs Our Teambuilding Props

Anyone who has ever prepared props for a corporate event knows the cycle. The idea takes a minute; then come hours of modeling in Tinkercad or another CAD tool, dialing in dimensions and exports. What if two sentences of description and a reference image were enough, and a few minutes later a finished 3D model opened in your browser? That is exactly what we built: a skill for Claude Code that takes a written brief, designs the product, generates a parametric 3D model, shows an interactive visualization, and exports an STL ready for your slicer.

What does the skill actually do?

A skill is a set of instructions and small tools that teach an AI assistant to act as a product designer and CAD operator in one. It takes a description as input, say “a 46 x 30 mm token with lettering and hex rivets in the corners”, or reference images. It returns:

• concept designs and variants as generated images for a quick decision,
• a parametric 3D model in OpenSCAD, or in Python via build123d when the shape needs rounding beyond what OpenSCAD can do,
• a check that the model is watertight and printable,
• an interactive 3D preview in the browser where you rotate the part and flip between front, back, and side views,
• an STL or 3MF export plus a filament estimate for the whole batch.

It all runs on top of Agentic OS, the template we use at Printara3D as the base for our automation. You need Claude Code and a handful of small open source tools; the skill walks you through installing them.

What does the path from brief to print look like?

A wizard first asks about purpose, dimensions, text, and material. Then it picks a route: an angular part with graphics goes to OpenSCAD, a more organic shape goes to build123d. The model gets rendered from several angles and the AI compares the result with the brief on its own; when text overflows an edge or elements collide, it adjusts the parameters and renders again.

That QA loop is the most valuable part of the whole approach. Dimensions are not guessed; they are measured from the first render. A concrete example from practice: the line width of the Rajdhani Bold typeface comes out at roughly 0.72 times the font size times the character count. The first estimate was 0.52, and the text collided with the corner rivets. Since then, the ratio gets measured, not guessed.

Use case: tokens for the Survival program

The skill’s first real job was game tokens, the in-game currency and rewards, for Survival, the new teambuilding program by Z-AGENCY. The brief: a rectangular token with “Project Zone” lettering, a denomination, and a coin-style notched edge so it handles well even with gloves on.

The resulting 10-value token measures 46 x 30 x 5 mm, has a volume of 6.12 cm3, and weighs about 7.6 grams of PLA. The two-sided relief is mirrored on identical coordinates, so the front and back register like a struck coin. The relief depth is a multiple of the layer height: 0.8 mm at a 0.2 mm layer. Bridges up to 6 mm print without supports.

The first batch: 40 pieces on a single plate, 17.5 hours of printing on a Bambu Lab H2S. From a preparation standpoint, the difference is fundamental. Modeling alone used to eat an afternoon; now that same time went into design iterations, the part that actually shapes the result.

Where are the limits? The rubber duck test

Honesty requires admitting where it creaks. The skill handles angular and rotational shapes reliably. But when it was asked to model a small yellow rubber duck, the iterations ran nonstop and the result still had a long way to go before you’d hand it to a kid in the bathtub. Organic shapes without reference geometry remain the weak spot of code-described modeling. For figurines and sculptures, sculpting or 3D scanning is still the better path.

Pros and cons of generated models

Pros:

• Speed from idea to STL is measured in minutes, not hours.
• A parametric source: changing the denomination or text means editing one variable, not remodeling.
• Repeatability: the second denomination came from copying the file and adjusting parameters, with no risk of regressing a part already in production.
• The material estimate for the whole batch is part of the output; 40 tokens come out to roughly 305 g of PLA.

Cons:

• Organic shapes are still out of reliable reach.
• Dimensions must be checked by comparing with the render, not trusted blindly on the first pass.
• For engineering-grade precision with tolerances, classic CAD remains irreplaceable.

Why connect AI with 3D printing at all?

The most interesting moment in all of this is when a digital idea stops being pixels on a screen and becomes an object jingling in a colleague’s pocket. AI can generate text and images in endless quantities, but 3D printing is one of the few paths that carries its output into the physical world without a factory and without a minimum order of a thousand units.

For a small company, that means having its own micro-manufacturing. The design comes together in the morning, printing runs in the afternoon, and by evening the team is playing with the tokens. The human role doesn’t shrink into spectating, either: a person writes the brief, decides between variants, and approves what goes on the plate. The AI does the craft work that used to swallow most of the time.

Try the skill yourself

You can download the exported version of the skill, in English, version 1, as an attachment below this article and install it into your own Claude Code. Inside you’ll find the complete instructions, scripts for both the OpenSCAD and build123d routes, the QA loop, and the 3D viewer. You need Claude Code, Python, and OpenSCAD; the skill installs the rest and walks you through your first model.

Conclusion

Generative part design works best today where the shapes are parametric, angular, or rotational, with text and graphics: tokens, tags, holders, bookmarks, simple mechanical parts. The key isn’t the generation itself but the measure-and-check loop that compares every render against the brief. And the limits are clear: the rubber duck will have to wait; 40 tokens overnight and change already don’t.