Skip to content

CheshireJCat/create-3d-model-skill / skills/create-3d-model

create-3d-model

Verified

Create and refine portable 3D models from text prompts, one or more reference images, wireframes, orthographic views, templates, texture atlases, or existing 3D assets. Use whenever Codex must model or reconstruct geometry, match visual references, create materials and UVs, rig or animate a model, validate shape or multiview fidelity, repair a subpar asset, edit an existing Blender scene, or export GLB, glTF, FBX, OBJ, USD, STL, or another requested 3D format. Uses Blender and BlenderMCP as the production backend while treating the transferable 3D model as the normal deliverable.

Create 3D Model

Turn text, images, wireframes, multiview references, or existing assets into validated 3D models. Use Blender as the production backend and editable source environment, not as the definition of the product. Inspect the inputs and open scene, route the request to the smallest relevant module set, execute controlled bpy changes, inspect visual evidence, iterate, export the model, and preserve the engineering source.

This Codex skill keeps the complete 30-part upstream capability stack inside one discoverable skill. Treat this top-level skill as the orchestrator and load the other 29 modules on demand from references/modules/.

Start here

  1. Read references/codex-integration.md before the first Blender MCP call in a task.
  2. Read references/capability-map.md and select only the modules needed for the request.
  3. When a module says to load or chain-load another skill, read references/modules/<skill-name>/SKILL.md. Resolve that module's relative links and scripts from its own directory.
  4. Treat this file and codex-integration.md as higher priority than a vendored module when tool naming, scene safety, output paths, self-modification, or Git guidance differs.

Do not read every module preemptively. Load the orchestrator or domain module first, then follow a dependency only when its gate is reached.

Discover Blender MCP capabilities

  • Find the connected Blender tools by semantic capability or suffix. Typical operations are get_scene_info, get_object_info, get_viewport_screenshot, and execute_blender_code; the exact namespace may vary.
  • If the Blender tools are unavailable, verify that Blender is open, the BlenderMCP add-on is listening on 127.0.0.1:9876, and the current Codex task started after MCP registration. Do not claim Blender work completed while the connection is unavailable.
  • Keep Poly Haven, Sketchfab, Hyper3D, Hunyuan3D, and other network asset or generation features disabled unless the user explicitly requests them. Never expose the Blender control port beyond localhost.

Follow the production workflow

1. Establish the target

  • Classify the input as text-only, single-image, multiview or orthographic, wireframe or technical drawing, texture or template pack, or an existing 3D asset. Combine modalities when the user supplies more than one.
  • Extract the subject, intended use, style, dimensions, reference priority, primary delivery format, target runtime, animation requirements, and quality bar.
  • Inspect every provided image before modeling. Record views, landmarks, silhouettes, part counts, material regions, uncertainties, and contradictions.
  • For a single image, state depth, scale, occlusion, and unseen-surface assumptions. For multiple images, register their view relationship before adding depth.
  • Use reversible defaults when missing information does not materially change the result. Ask one focused question only when a choice changes geometry, topology, animation, source-of-truth policy, or delivery.
  • Read references/core/common-object-dimensions.md for common subjects; prefer user-provided or authoritative dimensions.

2. Inspect before mutating

  • Inspect the scene before the first write. Inspect each object in scope and obtain a viewport screenshot when image capture is available.
  • Treat existing objects, collections, materials, lights, cameras, worlds, actions, and files as user-owned.
  • Do not clear the scene, reset the world, delete the default cube, or replace the camera merely because a recipe starts from a blank scene.
  • Delete startup content only when the user requested a fresh scene and inspection proves it is unchanged startup content. Otherwise create a task-specific collection and leave unrelated data alone.
  • Save a versioned checkpoint before broad or destructive work. Never overwrite the user's original .blend without explicit authorization.

3. Route and plan

  • For a complete scene or multi-stage asset, load blender-pro-workflow first, then the required domain modules.
  • For multi-module or source-locked work, load blender-skill-harmonizer so one module owns each decision and handoff.
  • For exact reference work, establish a source manifest and validation gates before creating geometry. Do not mix plausible freeform modeling with 1:1 reconstruction after fitting begins.
  • Define units, origin, up axis, long/broad/thin axes, object boundaries, stable names, collection ownership, modifier strategy, output paths, and validation views before executing code.
  • Prefix newly created data with GEO-, MAT-, LGT-, CAM-, ARM-, COL-, or WGT-.

4. Build from blockout to detail

  • Follow references/core/assembly-order.md for multi-phase work.
  • Establish dimensions, silhouette, camera framing, and part connectivity with the fewest primitives or contours needed.
  • Capture and inspect a blockout screenshot before adding secondary detail. Fix scale, orientation, silhouette, framing, disconnected parts, and source registration first.
  • Prefer Blender's data API or bmesh for deterministic edits. Use bpy.ops only with explicit active object, selection, mode, and context.
  • Treat every code-execution call as a fresh Python namespace. Re-import modules and retrieve persistent Blender data by stable name.
  • Split work into small, purposeful, retryable chunks. Prefer idempotent creation and update code; print compact structured evidence after each chunk.
  • On timeout or error, diagnose and retry only the failed chunk. Do not replay an entire scene build blindly.

5. Validate structure and appearance

  • After each significant phase, verify names, transforms, dimensions, origins, hierarchy, modifiers, topology counts, normals, UVs, materials, actions, accidental duplicates, and file paths relevant to the task.
  • Inspect a viewport image after blockout and final refinement. For look development, camera, lighting, render, texture, or animation work, inspect rendered evidence as well.
  • Use Codex's local image-viewing capability for saved renders. A successful tool call or non-empty image file is not visual validation.
  • For animation, render representative frames or a contact sheet and load animation-quality-gate.
  • For reference-locked work, use measurable masks, landmarks, overlays, bboxes, centroids, IoU/SSIM where appropriate, and multiview gates. A front-view match alone does not prove a 3D match.
  • If no image can be inspected, report structurally verified, visually unverified.
  • Identify the largest visible mismatch, change the smallest upstream cause, re-capture the same view, and compare before and after. Do not hide geometry errors with camera, lighting, or crop tricks.

6. Save and hand off

  • Treat the transferable 3D model as the primary artifact. Use the user's requested format; when none is specified, export a single-file .glb and save a versioned .blend engineering source.
  • Treat .blend as the editable project file, not the only proof that a 3D model was delivered. A preview is visual QA, not a substitute for the exported model.
  • Let explicit 3D-printing, engine, CAD-interchange, render-only, animation-only, or existing-project targets override the default .glb contract.
  • Write final artifacts to absolute paths in the active workspace or another user-approved durable directory, never /tmp or an expiring BlenderMCP screenshot path.
  • Verify every deliverable exists, is non-empty, and matches its requested format. Re-import or inspect the export when practical to catch missing geometry, textures, materials, axes, scale, or animation.
  • Report created or changed objects, important dimensions and topology, modules used, validation evidence, and remaining limitations.
  • End with an Artifacts section. Link each verified output with a descriptive Markdown label and its absolute local path so Codex can expose a quick-open file reference. List the primary model first, the versioned .blend source second, and previews or additional exports after them. List each path once and never fabricate an artifact link.

Run bundled helpers carefully

The upstream analysis and validation scripts remain beside their modules under references/modules/<module>/scripts/. Before running one:

  1. Read the module's SKILL.md and inspect the script's --help or source.
  2. Resolve the script from the installed skill directory and pass absolute input and output paths.
  3. Keep outputs in the active workspace unless the user approved another location.
  4. Use an existing environment with Pillow, OpenCV, NumPy, and SciPy, or run ephemerally with uv and only the dependencies the script imports.
  5. Treat generated reports and Blender recipes as evidence or plans. Review generated Python before sending it to Blender.

Do not run release, documentation, commit, or push steps from quality-refinement-autoloop unless the user explicitly asks to develop or publish the skill itself. For normal modeling, use that module only to diagnose the artifact and plan a scoped repair.

Enforce safety boundaries

  • Review downloaded, generated, or user-supplied Python before execution.
  • Do not enable global script auto-execution, install add-ons, download assets, call paid services, or change Blender preferences unless explicitly requested.
  • Do not perform scene-wide deletion, destructive remeshing, modifier application across unrelated objects, mass relinking, or overwrite saves without authorization.
  • Do not claim production human anatomy from primitive recipes. Use an authorized base mesh, sculpt workflow, or specialist when required.
  • Redirect engineering-grade CAD tolerances, simulation validity, and manufacturing certification to appropriate CAD or CAE workflows.

References

reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 468kUpdated Claude CodeCodexUnverifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 319kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 319kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 319kUpdated Claude CodeCodexVerifiedInstall