How to use Claude to design GPU fan shrouds
Claude generated GPU shrouds from caliper photos, then Andrew Leech used pinned 3D comments to correct geometry before printing and installing the final parts.
Original prompt
Shared by the source author. See source notes for context.
no that sounds wrong. I know text descriptions is hard, but that does not sound right at all.
Reach people exploring AI agents
See placements →What happened
Leech printed and installed the GPU shrouds and duct and published meshpin, a reusable coordinate-based model-review tool.
Reported by Andrew Leech on BlogHow the agent did it
- Leech photographed caliper readings from the physical heatsink.
- Claude generated CadQuery geometry and revised it from marked-up photographs.
- They built coordinate-based meshpin review, then Leech printed and installed the parts.
Where a person stepped in
Leech measured the physical heatsink, reviewed renders, annotated photos and printed and installed the final parts.
Andrew Leech designed printed fan shrouds for a dual NVIDIA V100 board with Claude Code. Because there was no CAD file for the second-hand heatsink, he supplied photographs of caliper measurements. Claude read the displays and incorporated dimensions into a Python-generated CadQuery model.
Geometry generation worked better than spatial discussion. Claude repeatedly misunderstood which face Leech meant. He replaced verbal corrections with marked-up photographs showing mounting ears, shroud placement, fan location and open versus sealed fins. The annotations let the agent map airflow to the model's axes.
Rendered PNGs introduced another failure: an old browser image made correct geometry appear wrong. The pair built meshpin, a local STL viewer that stored comments with actual three-dimensional coordinates beside the model files. Leech could pin a requested change to a precise face, while Claude added its own callouts for review.
The post shows the final shrouds and central fan duct printed and bolted onto the cards. It also links the reusable viewer. The concrete result is installed physical parts and a clearer review tool, while the account does not provide measured cooling performance or eliminate the need for human dimensional checks.
Prompt
Original prompt
Source: Blog · Andrew Leech
Frequently asked questions
How were dimensions captured?
Leech photographed digital-caliper readings and Claude entered them into the model dimensions.
Why replace PNG review?
Stale exports and browser caching made the displayed geometry unreliable.
Were parts actually printed?
The account shows the shrouds and central duct installed on the V100 board.
More details
About this account
Geometry generation worked better than spatial discussion. Claude repeatedly misunderstood which face Leech meant. He replaced verbal corrections with marked-up photographs showing mounting ears, shroud placement, fan location and open versus sealed fins. The annotations let the agent map airflow to the model's axes.
Primary account opened. Printed parts are pictured installed. Original prompt excerpt preserves author's correction; thermal performance is not measured in the post. Original prompt excerpt retained from the source.
Credit: Andrew Leech on Blog
Source published 23 Jul 2026
Source checked 14 Oct 2026
Tools and topics
Claude Code · CadQuery · Python · meshpin · 3D printer
GPU cooling · Parametric CAD
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.