gpt_3d_skill
A vendor-agnostic collection of skills for procedural 3D modeling, animation, interactive WebGL walkthroughs, and character rigging with local motion capture, with a reusable mobile viewer template.
Capabilities
- Procedural 3D Modeling (
blender_modeling): Deterministic geometry generation in Blender, procedural material assignment, structured scene hierarchies, geometry and normal checks, and self-contained asset packaging. - Cinematic Architectural Animation (
blender_animation): Multi-stage choreographies (orbit showcases, exploded architectural views, assembly sequences, staged object entry), matrix transform synchronization, and frame-accurate FFmpeg video encoding. - Blender Compositing Plate: Frame-accurate Blender layers for another renderer to composite over, with a shared coordinate system and per-frame camera file, per-frame projection checks, formula-matched lighting, and deterministic analytic cloth deformation shared by both renderers.
- Hybrid 3D / AI Video: Blender start frames with generative character acting, per-beat narration, unified scoring, foley alignment, and explicit editorial and QA boundaries.
- Interactive Web Walkthroughs (
web_walkthrough): Static high-fidelity WebGL delivery via texture baking, glTF/GLB export, pre-batch collision boundaries, desktop pointer-lock fallback, and responsive mobile touch navigation. - Character Modeling from Reference Sheets: A static toon character built from turnaround and detail sheets. Matched-view renders at the sheet's own scale, silhouette and limb measurement, part-wise generators suited to parallel agents, base-mesh faces with an image-model repaint projected back onto the head, and a web export budget. Includes a review template.
- Character Rigging and Motion Capture (
character_rigging_mocap): Reference-driven motion-model selection, rest binding, skinning and browser retargeting. Includes camera-free presets, local capture lifecycle handling, clean asset builds and separate rig, fixture, live-device and user-approval evidence. - Offline Model Sharing: A single HTML file containing the interactive model, viewer runtime and source-photo comparison. Recipients download it and open it in a desktop browser without a server or installed dependencies.
Ordinary model iterations start with a useful preview and checks relevant to the changes. Completed model handoffs include an offline comparison HTML by default unless the user requests another format. Extra renders and packaging are conditional, while requested production, animation, rigging and walkthrough checks remain in scope. See the router's iteration and handoff policy.
Installation
This repository provides loose Markdown skill definitions designed for direct integration by coding agents (such as Codex, Claude Code, Cursor, OpenCode) and human developers.
Reuse the Viewer
The walkthrough template below is for first-person exploration. For a model that a friend can open by double-clicking a single file, follow the standalone sharing workflow; opening a development index.html with adjacent ES modules and GLB fetches under file:// is not equivalent to an offline bundle.
Start from templates/mobile_walkthrough/ instead of rewriting the controls. Copy that directory into a task workspace, supply your baked GLB and collider file, run npm ci, then npm run dev. No runtime demo or user model is bundled.
Configure public/scene.json with asset paths, background, spawn point, walk bounds and overview pose. The template keeps touch/keyboard movement, collision, reset/overview and one baked-unlit rendering path. There are no renderer modes or broad camera/quality knobs. BASE_PATH supports a website subpath. Existing unit and browser tests remain.
Agent-Assisted Installation
You can provide this repository location to your coding agent and request installation:
- The installer agent inspects the target workspace agent configuration files (such as
AGENTS.md,CLAUDE.md, workspace router instructions, or skills index). - The agent registers a reference pointer exclusively to the root router skill:
- Root Skill:
skills/skill_gpt_3d.md
- Root Skill:
- The seven specialized domain skills remain local to this repository and are navigated dynamically via the root router:
blender_modeling.md,blender_animation.md,blender_compositing_plate.md,hybrid_ai_video.md,web_walkthrough.md,character_modeling.md, andcharacter_rigging_mocap.md.
Give the agent the repository's GitHub URL or local checkout path. It should start from the target workspace's AGENTS.md or CLAUDE.md, follow WORKSPACE.md if present, and update rules/skills/INDEX.md or skills/INDEX.md. If no index exists, add one short root-skill pointer to the agent instructions instead. Register only the router, not every focused skill globally.
Local Usage
To use these skills locally, reference the root router:
- Root router:
skills/skill_gpt_3d.md
When invoking modeling, animation, walkthrough, rigging or motion-capture tasks, direct your assistant to consult skills/skill_gpt_3d.md. The router identifies task requirements, enforces workspace boundaries, and dispatches the appropriate specialized skill. Deforming character assets use the rigging branch rather than inheriting static-environment baking assumptions; image-to-video acting remains in the hybrid-video branch.
Artifact and Workspace Boundaries
User models, references, renders, deployment settings and builds belong in task workspaces outside this repository. The repository includes reusable template source and code-generated test fixtures, but generated browser outputs and dependencies remain ignored. No analytics or remote asset CDN is enabled by default.
Verification
Repository integrity is verified using standard Python unit testing:
uv venv
source .venv/bin/activate
python -m unittest discover -s tests -vThese tests validate Markdown links, document structure and repository hygiene, including template source. Run the template's npm test and npm run test:e2e for configuration, collision and browser checks. CI runs all three suites. They do not certify arbitrary user models or physical-phone performance.