All insights · Sergey Kosenkov — portfolio
Claude Code as a Motion Designer: Vibe Coding a Techno VJ Loop with Blender, Tesseract & After Effects
An honest production breakdown of human-agent symbiosis: from GPT Image 2 concepts and Sketchfab assets to Blender EEVEE, Tesseract WGSL Voronoi glass, and manual AE timing synchronization.
Sergey Kosenkov, Virtual Production (ICVFX) Specialist & AI Technologist · · 6 min read
Watch the video on YouTube
Concept Art in GPT Image 2 & Sketchfab Asset Sourcing
Establishing the cyber-organic aesthetic and sourcing clean base geometry under CC-BY-4.0
Stage 1: Cyber-Organic Concepting in GPT Image 2
You have likely seen videos circulating online where creators "vibe code" rave concert visuals using Claude acting directly inside Blender. The classic viral formula features a cybernetic skull rotating to intense techno drops amidst flashing HUD displays and kinetic typography. I set out to reproduce and elevate this experiment with an authentic, multi-tool agentic pipeline.
The journey began in ChatGPT with GPT Image 2 to establish a visual benchmark. I directed the model to design an intricate, high-contrast gothic-cyberpunk bust: an organic weathered bone skull fused with titanium cranial plates, glowing red ocular diodes, hydraulic cheek pistons, and intertwining crimson roses studded with cybernetic thorns.
💡 Key Strategy: Instead of vague prompt iterations, feeding Claude a high-fidelity visual concept from GPT Image 2 gave the agent a clear geometric blueprint and color palette to replicate across the entire pipeline.
Initial cyber-organic concept generated in GPT Image 2 establishing the visual benchmark.
Stage 2: Intelligent Asset Sourcing on Sketchfab & Model Attributions
In the original viral experiment, the author relied on Tripo3D to generate a raw skull mesh. However, AI text-to-3D models often yield unified, "melted" meshes that lack clean topology and separate articulation points needed for mechanical animation.
To solve this, I instructed Claude Opus 5.5 to browse Sketchfab directly, search for modular assets, and select high-quality base components under Creative Commons. The agent identified and downloaded an anatomically authentic skull with a separate movable jaw, alongside several high-res 3D rose models with distinct stems and petal geometry.
⚖️ 3D Assets Attribution (CC-BY-4.0 License):
• "Skull downloadable" by martinjario
• "Red rose" by Lassi Kaukonen (thesidekick)
• "Rose" by Heliona
All 3D assets sourced from sketchfab.com under Creative Commons Attribution 4.0 International.
Base skull asset sourced autonomously from Sketchfab ("Skull downloadable" by martinjario).
High-res 3D rose assets with separate leaf and thorn structures (by Lassi Kaukonen and Heliona).
Autonomous 3D Assembly, Rigging & EEVEE Dual-Pass Shading
Directing Claude via Blender MCP to slice meshes, install mechanical hardware, and configure EEVEE render passes
Stage 3: Robotic Dissection & Mechanical Kitbashing in Blender
With the base assets imported into Blender via the Model Context Protocol (MCP), I tasked Claude with dissecting the skull and fabricating mechanical upgrades inspired by our GPT Image 2 reference.
Working programmatically via Python scripts, Claude sliced the cranium to install structural plates on the forehead, mounted cylindrical "headset" housings at the temples, fitted pneumatic pistons along the cheekbones, added a stabilizing bracket to the lower jaw, routed flexible conduit hoses along the neck, and wove thorny metallic stems throughout the bone cavities.
Crucially, Claude kept every mechanical component on individual transformation hierarchies. This allowed the plates, pistons, and roses to oscillate, pulse, and rotate independently in sync with camera motion.
⚙️ Director's Iterations:
• Layer G3 was moved behind the skull with beat-synced strobing to avoid obstructing foreground laser sweeps.
• The .ART logo element was programmed to start blinking on the rhythm switch at 4.0s.
• Stepped typographic layout (Ai / UNREAL / .ART) aligned along the top angle of the roses.
Complete assembled cyborg skull bust in Blender with procedural mechanical kitbashing.
Stage 4: Blender EEVEE Shading & Dual-Pass Output (Hero + Glow)
To maintain rapid iteration speeds and avoid heavy offline render times, the scene was rendered in Blender EEVEE (not Cycles and without Subsurface Scattering). Shaders were built with high-contrast surfaces: an aged bone material featuring a clearcoat layer, dark brushed metal plates, chrome hydraulic rods, and intense emissive shaders for the ocular LEDs and headset rings.
In the Blender compositor, a Bloom pass was applied to give the red LEDs organic atmosphere. The scene was exported as two distinct 1080p@30fps image sequences: a hero RGBA beauty pass and a dedicated isolated glow pass for downstream optical compositing.
🎯 Render Specs: 1920×1080 @ 30 fps (1200 frames total, 40s duration). Blender EEVEE with dual-pass output: Hero RGBA + dedicated Glow pass.
Rhythm Mapping, Tesseract 2D HUD & Manual AE Timing Integration
The true hybrid pipeline: Tesseract CLI layout ➔ After Effects manual timing ➔ Tesseract final render
Stage 5: Suno Rhythm Analysis & Exact Numerical Mapping
While the agent was constructing the 3D model, I generated an industrial techno loop in Suno. Once exported, I selected a 40-second slice (from 0:55 to 1:35, exactly 1200 frames at 30 fps) and fed the audio file to Claude for rhythm breakdown.
Claude calculated the exact tempo of 140 BPM and identified key transient drop intervals. Rather than relying on guesswork, the animation curves were mathematically locked to concrete timestamps: the bar 8 crack impacts were timed precisely at 12.05s, 12.91s, and 13.33s, leading directly into the explosive drop at 13.763s.
⏱️ Concrete Rhythm Coordinates: 140 BPM; 40s duration (1200 frames @ 30 fps); Bar 8 impacts: 12.05s / 12.91s / 13.33s; Main drop: 13.763s.
Stage 6: The True Pipeline: Tesseract CLI ➔ After Effects ➔ Tesseract Final Render
A common misconception about agentic workflows is that After Effects is where final renders happen. In reality, our architecture was far more sophisticated: Tesseract (via Tesseract CLI + agent skills) initially generated the entire 2D motion HUD layer.
A custom Python bridge script (to_ae.py) translated the Tesseract document into Adobe After Effects, converting precomps, shape layers, and HUD animators into editable native expressions. From there, I manually nudged word layers and typography timings on the AE timeline to achieve perfect artistic feel.
The dynamic HUD tracks song lyrics in real-time through AE expressions: the reticle crosshair locks target on the word UNREAL, status indicators display the active line, and AGENT: ON activates exactly on the vocal hit "ON". When I moved word layers on the timeline, the HUD automatically re-aligned itself. Once timings were locked, export_timings.jsx dumped the refined timestamps into JSON, and Tesseract rebuilt and rendered the final 2D document.
🤝 Human-Agent Symbiosis: "The human director moved layers on the timeline to taste, while the agent reprogrammed the HUD, dynamic Voronoi glass, and typography to follow suit."
After Effects composition showing SKULL_VJ_MAIN with Claude’s scripted layer groups, markers, expressions, and Blender passes.
Stage 7: GPU Voronoi Glass Shatter in Tesseract (Custom WGSL Shader)
Instead of a traditional 3D simulation or proprietary RTX effect, the shattered glass was authored as a custom WGSL Voronoi cell shader running directly on the GPU inside Tesseract.
The Voronoi crack network propagates sequentially on the three acoustic hits of bar 8 (12.05s, 12.91s, 13.33s), accumulating mechanical tension before completely shattering into refractive shards at the drop timestamp of 13.763s.
16-shot contact sheet showing camera moves, typographic cuts, and lighting.
Dual-Agent Review with Fable 5.2, Production Realities & Cover
Cross-model peer review, debugging pipeline roadblocks, and programmatic Python cover assembly
Stage 8: Fresh-Eyes Review with Claude Fable 5.2
To avoid context fatigue, the final review was conducted in a fresh, isolated session using claude-fable-5-2 with a concise technical brief and zero prior chat baggage. This "fresh pair of eyes" technique allowed the senior model to critique the cut objectively.
According to NOTES.md, Fable 5.2 implemented crucial directorial upgrades: establishing the typographic branding "Ai UNREAL .ART", tightening word impacts (0.2s rise, 2-frame white impact flash, chromatic cyan and purple ghosts collapsing into the letters), adding shockwave blast rings on drops, generating a background text wall behind the skull, and shortening the drop flash so the shattering glass remains crisp.
In audio, Fable redesigned the ending with a reverse-swell, punchy kick, sub, and crash on the final beat, while normalizing master loudness to approximately -14 LUFS (the standard for YouTube streaming delivery).
What Went Wrong: Roadblocks & Realities in Desktop DCC Tools
No production pipeline runs without friction, and discussing the failures is far more valuable than marketing hype:
• Font Installation: After Effects refused to recognize fonts installed only for the current user profile. We had to reinstall fonts system-wide for all users and strip static font family variants to prevent script crashes.
• aerender Hangs: Automated background rendering via aerender hung repeatedly whenever the main After Effects GUI application was running simultaneously.
• Semi-Manual Script Triggers: While Claude wrote to_ae.py and export_timings.jsx, executing them inside AE required manual human operator triggers.
⚠️ Practical Takeaway: AI agents excel at generating complex scripts and mathematical coordinates, but desktop DCC software still requires human oversight to navigate OS permissions, font registries, and process locks.
Production Efficiency & Programmatic Python Cover Assembly
The total turnaround time for the project was approximately 3 to 4 hours, consuming 60–70% of a standard 5-hour Claude subscription limit in High reasoning mode.
The official cover art was assembled programmatically by Claude using a custom Python script that layered the hero render, glow pass, and Voronoi glass shard frame with graphic overlays—no text-to-image generative models were used for the cover.
📊 Production Metrics: 3–4 hours turnaround; 60–70% of 5h High mode limit; 1080p @ 30 fps master delivery (1200 frames); Zero manual vertex modeling.
Official cover art assembled programmatically by Claude in Python from hero, glow, and Voronoi glass render passes.
Agentic Motion Design: Lessons from the Trench
Key takeaways from bridging AI coding agents with professional motion graphics software
🚀 The Reality of Agentic Motion Design
This project demonstrates that AI coding agents like Claude can act as capable technical directors and motion design assistants when connected to specialized tools through scripting and CLI bridges. Rather than replacing human artistry, the agent excels at tedious mathematical calculations (BPM-to-frame mapping), procedural rigging, and script generation, while the human director provides aesthetic taste, manual timing tweaks, and creative direction.
⚡ Key Pipeline Insights & Asset Credits:
- Curated 3D Topology: Sourcing clean CC-BY-4.0 assets from Sketchfab avoids the melted geometry of current text-to-3D models.
- Mathematical Musical Sync: 140 BPM, drop at 13.763s, and bar 8 Voronoi cracks ensure tight audio-visual cohesion.
- Tesseract ➔ AE ➔ Tesseract Loop: Using AE as an interactive timing editor while Tesseract handles GPU WGSL rendering.
- Fresh-Eyes Model Handover: Spawning a clean session with claude-fable-5-2 for unbiased final polish and -14 LUFS mastering.
- 3D Asset Credits (CC-BY-4.0): "Skull downloadable" by martinjario · "Red rose" by Lassi Kaukonen (thesidekick) · "Rose" by Heliona — sketchfab.com