Skip to content
← Articles

November 06, 2026·3 min read

A skill that picks between Three.js, Blender and AI video, and keeps itself honest

I built the same scene two ways and turned what I learned into an open-source Claude Code skill: it picks the tool by the job, gives the render and assembly recipes, and can tell what was measured from what is only research.

After building Ezekiel's wheels two ways, the question left over was not about the scene. It was: next time, which tool do I use, and why? Three.js, Blender or an AI video model?

I had two measured numbers and a third option I had never run. Instead of deciding from memory on every project, I wrote the decision down as a Claude Code skill and open-sourced it: video-3d-skill, MIT.

The choice in 38 seconds. The AI-video part is a card: I generated no footage with it.

What it decides

The skill starts with a short brief: what the job is, whether the geometry must match a spec, whether the output must be interactive, what the budget and deadline are, and how much the audience will hold you to honesty about AI use. Then it picks:

  • Three.js (with the teardown skill), to explain a mechanism exactly and interactively. Measured: 36 min of build, 82 s of capture for a 16 s film.
  • Blender with Cycles, for a cinematic shot that must match a spec. Measured: 504 frames, 4.13 h of render, about 29 s per frame on an Apple M4.
  • AI video, for atmosphere (sky, fire, smoke), plausible motion and a fast first draft, and only with the cost approved.
  • Hybrids, ranked by how well each is documented. What I did was the simplest one: the console explains and the Blender scene convinces, cut together into one film.

The rule of thumb, with the evidence in the repo: AI video is weak at exact geometry, counting and text in frame; Blender is weak at organic motion and some materials; Three.js is cheap and exact but looks like a console, not cinema.

Measured and researched are not the same thing

This is the part of the skill that matters most. The Three.js and Blender parts come from a real project, with metrics in a file the repo copies into evidence/. The AI-video part is research dated 2026-11-06, and I have never run a video model. Prices come from resellers and drift, and the skill says so.

I got this wrong myself, which is why the skill got stricter. The agent claimed my API key only had image models, and I published that line in the previous post. It was false: my search filter looked for "veo", "imagen", "image" and "video" and missed the gemini-omni models, which generate video. I found out later, fixed the post, and the skill now tells you to filter for omni too and to mark the result as untested.

The rules it enforces

  • Never call a paid generation endpoint without your approval of the cost. Listing models is free; generating is not.
  • One Blender process at a time. Two at once hung both on my Mac.
  • Render in 8-frame chunks, resumable, with a watchdog. One chunk hung under memory pressure, with swap at 9 GB, and cost about 50 minutes. The repo has the generic script I now use: it skips finished chunks, refuses to start if another Blender is running and restarts a stalled chunk.
  • Log what looks wrong. Time, attempts, seconds per frame and an honest list of flaws. When a number was not logged, the skill writes "not logged" instead of guessing.

How to use it

git clone https://github.com/rwspatin/video-3d-skill.git
ln -s "$PWD/video-3d-skill/skill" ~/.claude/skills/video-3d-production

SKILL.md is the router, and the recipes live in references/: headless Blender, Three.js, film assembly, metrics and AI video. templates/ has the chunked render script and a captions assembly example.

If you try the AI-video path, with the cost approved first, tell me what worked. It is the part that most needs real data, and I take PRs.

Credits

  • The teardown skill by elvarofficial (MIT)

  • Skies, terrain and textures: Poly Haven (CC0)

  • Scripture in the examples: Douay-Rheims (public domain)

  • Renan