How this library was built

The 3D library on the home page was built by AI agents working to my specs, and this page records how, with a source for every claim.

Ongoing, last updated 23 October 2026

The home page of this site opens into a small 3D library. Each of my builds is an exhibit in the reading room, my papers are real books on the shelves, and every exhibit links to a plain page like this one. I did not write its code. I set what it had to do, made the design calls, and approved the result. AI agents did the rest, and this page records who did what.

Who did what

The first build ran in four steps, all on 22 October 2026. The AI-made furniture came after it, on 22 and 23 October.

  1. Design. In a design session, Claude questioned me one point at a time until the plan was settled: a library as the setting, a cut-away overview you can click into, live exhibits, and a plain page that every reader can fall back to. The answers became a written specification.
  2. Art style. Claude built three throwaway versions of the same room: warm and lamp-lit with simple geometry, a painterly cartoon style, and a realistic one. I picked the first. The painterly filter made the newspaper headline and the book titles unreadable, and the realistic room was the heaviest to download.
  3. Build. Claude turned the specification into a build spec of about 3,300 words with numbered checks, and Codex, OpenAI's coding agent, wrote the code. Claude then re-ran every check itself in a sandbox and had a separate reviewer read the code cold. The reviewer found eight problems worth fixing, none of them blocking. Codex fixed them in one more round.
  4. Approval. I reviewed a preview and this page before anything went live.

What had to be true

The library is an extra, never a gate. The home page shows my name and my work as ordinary text from the first moment, and the library takes over only if it is ready within 4 seconds. Otherwise the page stays plain and offers a button. It also stays plain on phones, on devices without WebGL2, and for anyone whose system asks for reduced motion; if they do enter, the camera cuts between exhibits instead of flying.

The whole library costs about 435 KB to download, against a budget of 5 MB, and the room is drawn in 46 draw calls. In a test that slows the network to 20 Mbps and the processor to half speed, it was ready in under a second and a half, even with the graphics drawn in software rather than on a graphics card.

Where the exhibits get their data

If a live source fails, its exhibit shows the most recent saved copy and says so.

The furniture the AI made

The first build had no AI-made objects. TripoSR, the one 3D model that ran, turned the newspaper stand into something like a crumpled paper bag, and TRELLIS, a stronger model, could not be reached. So the first room was built entirely in code from simple shapes.

On 22 and 23 October Claude tried again, running TRELLIS through my Hugging Face account. For each of the four large exhibits it wrote a description, had an image model, Qwen-Image, draw it in the room's style, and had TRELLIS turn that picture into a 3D model. The Wyrmling exhibit started from my own picture instead, a black feathered wyrm writing in a book, which I made with ChatGPT; Qwen-Image-Edit redrew it in the room's style before TRELLIS. I compared each model with the code-built one from the same camera, and picked the generated one all four times.

None of them went in untouched:

For each object, the provenance file lists the prompts, the model versions and licences, and every fix.

Where the record is

Every piece of code Codex wrote is tied to the spec it was built from by a hash, recorded in the commit and in a ledger in the site's repository. Claude wrote the code that fits the AI-made furniture into the room, and its commits say so. Building this site took three Codex runs, one for the plain site and two for the library, and 497,543 Codex tokens.