Claude Tips for 3D Work

Claude tips for 3D work

March 13, 2026

Although Iâm sometimes uneasy about it, Claude Code is an invaluable tool for my current projects. I still occasionally hand write code in NeoVim on the bits I care the most about (CSS, design and early architecture like API patterns), but I now leave most of the actual code writing to the machine once I have the skeleton up. Thatâs not to say I think the code it outputs is always great. I perform a heavy dose of sanitation after the fact, though even that is done through more suggestions to Claude. I do watch the edit feed live, and have built up the skill of speed reading so that I can stop Claude in its tracks when its does something naughty.

This works for most web projects really well. Where Iâve found Claude still canât reason is in âvisualâ work that needs eyes. It can reason about CSS fairly well, and does an OK job understanding design language and color theory when prompted, but itâs the more complex tasks like spatial analysis in 3D where it always fails. Claude has great tooling for reading plain 2D images though, and when presented a set of images of 3D space from different angles, along with the right reference language, it can usually figure something out. A lot of my early work in Table Slayer, and Counter Slayer, both heavy 3D web apps, involved me manually taking screenshots and saying something like âcheck this out, you did it wrongâ. It took HOURS of back and forth mostly because I was the one who needed to feed it screenshots.

Using Claude to build complex 3D scenes like this can be tricky

Like most, I learned quickly that when working with Claude, you need to think in terms of âreadable outputsâ when working with LLMs. For most, this is a test layer. During the exploratory experience of design though, you typically donât have tests, and make do with asking it to create lots of temporary logging. For 3D projects this might be things like positioning and coordinates. This works to a degree, but is laborious and often done after youâve hit a wall of failure. CAD systems have lots of unions and subtractions. Just because thereâs a box there, it doesnât mean its visible.

Natively, Claude canât read an STL. It will tell you it can, but after you realize itâs just bullshiting you about the contents of the binary, youâll end up down a road where it wants to install a bunch of Python libraries to view the files. This is sort of useful for reading STLs, and will give it a way to rotate a camera and zoom in on certain positions, but its not very useful for a Three.js web app where your problem is in the CAD system building the 3D in the first place.

To solve those sorts of problem you need to generate a real context loop where it can read the state of the application, navigate the camera, add some debug markers and iterate. The core loop is asking it to navigate the app on its own, control the camera, and add visual markers (think red spheres to give it position context) to test itself. Iâve found generating these types of âiterative validationâ loops essentially for working with Claude. This allows it to double check its work before before it assumes it solves a problem.

Hereâs the loop for Counter Slayer.

The full Claude.md is here and goes into more detail, but this is the heart of the workflow. âThe loopâ gives it a way to navigate the app with playwrite, select different 3D objects (like trays or boxes), change the camera, place a red sphere on it at a certain position, check a screenshot, and see if that screenshot matches the request. More importantly it does it without asking me for confirmation.

As an example, hereâs the tooling to let it navigate around.

Overall Iâve started to change my thinking about how to work with Claude. Instead of expecting it to understand my requests, I almost always build tooling first to give us a shared language to discuss the project. Screenshot loops are a great way to build that language.