Motion & 3D · Intermediate
A towering hydraulic press in a sunlit workshop, a conveyor feeding it six things, and one unbroken camera move that loops forever. The crayons mushroom and snap; a glass bottle bursts in extreme slow motion while the camera circles it; a superball squashes and tears; a stack of plates shatters plate by plate; a red-hot billet is forged, its scale crazing and flaking off with sparks. Then a tungsten cube: the press stalls, judders harder and harder, a relief line jets oil and blows, the pressure lets go and the cube sits there unmarked, a glint along its edge, as the camera pulls back to the opening frame. Everything is Python that Claude Code wrote for Blender: the workshop, the press, the objects, our own Voronoi fracture and physics, a speed-ramped clock, an endless-line trick that makes the last frame flow into the first, the one-take camera and the Cycles render. You get four exact prompts with a check for each, a loop test, and the complete source.
The recipe
Every step gives you the action, the exact input (a prompt, settings or a file) and a check to compare your result against the original.
Claude Code — Claude Opus 5.5 (claude-opus-5-5) — wrote and ran every script — your Claude plan Blender — 5.2.2 LTS (Linux x64 tarball, run headless with --factory-startup) — modelling, the per-frame handler, Cycles, compositor glare — free, GPL Cycles — built into Blender 5.2.2, CUDA backend — path tracing on the GPU (OptiX failed to initialise under WSL2, CUDA works) — free Python — Blender's bundled 3.13 for the scene; system python3 with numpy, SciPy and Pillow — the physics, the clock, the camera checks, the loop test — free FFmpeg — 4.4.2 — master and Reels encodes, grain and vignette, contact sheet — free GPU — NVIDIA RTX 3080 Ti, 12 GB, in WSL2 on Windows — Cycles — your hardware The final render needs about 3 GB of GPU memory per Blender process. Any CUDA card with 6 GB or more should work. Provenance: Recorded October 6, 2026: Claude Code with Claude Opus 5.5 (claude-opus-5-5); Blender 5.2.2 LTS (Linux x64, headless, --factory-startup); Cycles on an RTX 3080 Ti with CUDA under WSL2 at 1080×1920, 30 fps, 25.0 s (750 frames), 64 max samples adaptive (threshold 0.025) with OpenImageDenoise, time-lapse shutter; render 8 hours 27 minutes (two processes, 16.7 process-hours, on a GPU shared with another job). FFmpeg 4.4.2: master 71.6 MB (CRF 15), Reels copy 39.3 MB (CRF 18, aq-mode 3, maxrate 30M, faststart), no audio. Loop test (blur sigma 2, SSIM): 0.9659 across the loop point on the frames, between its neighbours 0.9243 and 0.9785 (mean difference 2.98, between 5.12 and 1.89); 0.9647 on the Reels copy (neighbours 0.9267 and 0.9779). No external images, footage, textures, HDRIs, models or generative-media models were used.