On October 30, 2026 I told Claude to go all out on a wiki design, and it rendered the images on my graphics card without asking me first. When the first round landed, I typed this back: “Oh My Claude!!! What have you done?? These are AMAZING!!!”
The site is geekbar.wiki, a reference site about Geek Bar vapes for adults 21 and over. It has no store and no shop links, and it isn’t finished as I write this. Claude’s first job that day was the topical map: 191 pages across 11 silos, done in 27 minutes. The design came next.
The design brief I gave Claude
The design brief I gave Claude pointed at aiseo.gg, my Discord server’s website, and asked for its effort without its look. My exact words were “start the designs using the same effort that you put into aiseo.gg, but dont make them look like any part of aiseo.gg, just the effort.”
That was the whole brief. I said nothing about 3D, renders or hardware. Eighty minutes later I had four separate looks, each with a home page and a flavor page on desktop and phone.
| Time (MT) | What happened |
|---|---|
| 12:11 PM | I ask for the topical map |
| 12:38 PM | The map is done |
| 12:43 PM | I send the design brief |
| 2:03 PM | Four looks arrive: a comparison sheet, 20 screenshots and clickable pages |
| 3:05 PM | I send my reaction |
| 3:32 PM | I ask how it got to my graphics card |
What Claude built instead of generating pictures
What Claude built instead of generating pictures was a 3D model of the device, made in three.js from published dimensions (28.5 by 57.6 by 91.7 mm) and the maker’s specs. It left the logos off. Then it path traced the model, which means calculating how light bounces through a scene, so the glass and the shell read like a studio product photo.
No AI image generator touched these, and nothing came from geekbar.com. One model turned into four different sites. One look pulled the device apart down the page, with each part as a link. One turned renders into banknote-style line engraving. One arranged all 87 flavors into a wheel. The one I picked writes every fact as a plain sentence, with two devices lying side by side as the equals sign.
One model, four different sites, no AI images.
The look I picked also shows why this scales. Change the colors and re-render, and a flavor has its own picture in about 30 seconds.
What Claude said about my graphics card
What Claude said about my graphics card is why I’m writing this up. Here is the reply to my “AMAZING” message: “That’s the best thing I could hear, Bike! Your graphics card did a lot of the work: every picture comes from a 3D device I built and rendered on your RTX 3080, so nothing is AI-made or taken from Geek Bar.”
That is generous. The card did the light math, but Claude decided to build a model, built it, lit it, and fixed it when it broke. I’ve written before about what Claude costs me to build tools. This job was different, because the heavy lifting ran on hardware I already owned.
How Claude reached my graphics card
How Claude reached my graphics card was my next question, and the answer was plain. Claude Code can run commands on my PC, and that was all it needed. No special connection, no admin access. In its own words, “I didn’t ask you first.”
It told me it took four steps:
- Ran NVIDIA’s standard status tool, which reported the RTX 3080.
- Started a hidden copy of Chromium with settings that tell it to use the graphics card, separate from my own Chrome.
- Installed three.js and three-gpu-pathtracer inside the project folder only, 42 MB in all.
- Had that browser work through each picture through WebGL, the same feature 3D websites use, and saved the images to the project.
Claude says my drivers and Windows settings didn’t change and nothing was uploaded to do the rendering. It did turn off the GPU watchdog, a safety timer for stuck graphics work, in those hidden browsers only. The card ran near 100% during renders.
If you run a coding agent on your own machine, your hardware is within its reach too. That worked in my favor here. It also means the agent decides how hard your PC works unless you tell it otherwise.
Where the render job broke
Where the render job broke was at full size, when the big images started coming out black. Claude’s first read was the browser losing the card at that size, and that held up. It had been sending work faster than the card could finish, the queue piled up, and the connection died while the log still said done.
The fix was small: read pixels back after every draw, which forces the code to wait for the card. After that, the full batch was clean. Claude also spotted a modeling flaw in the first sharp render, thin plates sticking out past the rounded shell, and fixed it.
| Render job | Time on the RTX 3080 |
|---|---|
| Test hero frame | About 15 seconds |
| Main hero at full quality | About 80 seconds |
| Full batch for four looks | About 20 minutes |
| One flavor in the second round | About 30 seconds |
Why one render per page matters for SEO
Why one render per page matters for SEO, in my read, is that a big reference site can now have an image made for each page instead of a shared stock shot. Google says it pulls an image’s subject from the page around it, including captions and titles, so a picture built for that page has its text beside it.
I can’t show you a ranking result, because the site isn’t live. The limits are real too. The device is close to the real one, not an exact copy. In the round I picked, every flavor used the same model, even flavors sold on other devices, and the colors came from Claude’s reading of each taste. The design’s footer labels the renders as illustrations, and the design puts “For adults 21+” at the top of every page.
I asked for effort and got a studio. Claude gave my graphics card a share of the credit, and I’ll take that.
Sources
- three-gpu-pathtracer (GitHub)
- three.js
- WebGL API (MDN)
- Google Image SEO best practices (Google Search Central)