ArtisanCAD Agent Generates Industrial CAD Models from Language Prompts.

Generating complex industrial computer-aided design (CAD) models from natural language remains a significant challenge for existing text-to-CAD methods. These methods often struggle with vague user prompts, underspecified design intent, and the need for robust feature dependencies and production-grade geometry. A new system, ArtisanCAD, aims to overcome these limitations by integrating expert procedural knowledge. ArtisanCAD introduces a core component called CAD intermediate representation (CAD-IR). This executable procedural representation meticulously encodes parameters, ordered operations, tool bindings, dependencies, and verification rules. CAD-IR serves a dual purpose: it acts as a vehicle for distilling expert CAD procedures, such as CATIA operation recordings, into reusable, parameterized skills, and it provides a structured scaffold to transform high-level or ambiguous prompts into complete, executable CAD operations. The agent retrieves these expert-derived skills, instantiates and refines the CAD-IR, and then executes the procedure via a dedicated CATIA-MCP backend. It further employs multi-view visual feedback for iterative refinement, ultimately generating production-ready B-Rep models. Initial evaluations show ArtisanCAD significantly improves generation from intermediate prompts and can distill expert CATIA recordings into reusable skills for new design variants, demonstrating its potential for industrial applications.