Show HN: AI in SolidWorks

Perceived capabilities and limitations of LLMs for CAD

• Out-of-the-box models are described as “not great” at CAD and especially weak at true 3D/spatial reasoning (e.g., choosing wrong sketch planes, extrude vs. revolve, sweep path/section placement).
• They often succeed at basic shapes and reasonable dimensions (e.g., mug example with mm units) but struggle when precise, interdependent constraints are needed.
• Multiple users report poor results using LLMs with OpenSCAD for anything beyond simple parts (gears, molds, complex rounded shapes), often reverting to traditional CAD.
• There’s skepticism that general-purpose LLMs can reliably follow detailed specs or datasheets yet; current performance on technical documents is called “awful.”

Integration patterns and technical approaches

• Strong theme: don’t let LLMs talk directly to complex CAD APIs (SolidWorks C#/VBA, etc.); generated code frequently fails.
• More successful pattern: build a high-level CLI / query language and a plugin that translates structured commands (JSON/DSL) into CAD/EDA API calls.
• Emphasis on relative/geometric relationships (“left of”, offsets, clearances, design rules) and parametrization, while minimizing the model’s direct handling of raw numbers.
• Systems use:
  • Scene/feature queries (“closest distance between surfaces X and Y”),
  • Measurement/DRC sanity checks,
  • Windows UI Automation to drive GUIs and discover formats,
  • Direct manipulation of binary file formats (e.g., Altium) to avoid OS lock-in.
• Multi-model orchestration is common (e.g., one model for planning/actions, another for visual understanding or translating natural language to DSL).

Interfaces: text vs traditional CAD vs alternatives

• Debate over whether chat is the right UI:
  • Some argue existing parametric CAD UIs (especially SolidWorks) are near “final form” for precise work.
  • Others see text as empowering for non-CAD users who “only know English” and can iterate conversationally.
• Suggested hybrids:
  • Text plus visual suggestions/animations of possible operations,
  • Query/measurement questions (“how far is that hole from the edge?”) with interactive controls,
  • VR/gesture-based interfaces for more intuitive spatial input.

Who benefits and for which tools

• Perception that:
  • Professional engineers tend to use SolidWorks/CATIA/Altium and are fast enough that AI may add less value in basic modeling.
  • Hobbyists and occasional users on Fusion 360 / FreeCAD / OpenSCAD might benefit more from text-to-model, but may be less willing to pay.
• Requests for support beyond SolidWorks: Fusion 360, Rhino, AutoCAD/Civil 3D, and web-based tools.

Reliability, specialization, and future directions

• Concerns about business defensibility: when it’s easy to script desktop apps via LLMs, many can roll their own; some prefer local agents over hosted SaaS.
• Several participants think domain-specific or fine-tuned models (for CAD/PCB/Brep generation) will outperform generic LLMs; projects like SGS-1, flux.ai, and PCB layout systems are cited as examples of this direction.
• Opus 4.5 and newer models are perceived as noticeably better at structured graphics (SVG) and some CAD-adjacent tasks, suggesting room for rapid improvement.

Reactions to SolidWorks and to AI assistance

• Split views on SolidWorks:
  • Beginners find it extremely unintuitive, convention-heavy, and poorly documented for new versions.
  • Experienced users argue it’s among the most usable pro CAD tools; the difficulty comes from domain complexity and the need for serious practice or formal training.
• Emotional reactions to AI:
  • Some enjoy offloading tedious but necessary work (pin labeling, design rules, repetitive modeling).
  • Others feel a sense of loss as AI encroaches on the “relaxing, fun, craft” aspects of CAD and engineering work.