Blender Modeling Workflow

Blender Modeling Workflow

Read the shared execution guidance once per task for autonomous decisions, evidence cadence, and long-task continuity.

Create the asset as source-controlled Python plus generated .blend and .glb outputs. Treat the script as the durable source and Blender as the execution runtime.

For a render-only scene, route final delivery through the rendering workflow: source, self-contained .blend, images and render manifest are sufficient. Require GLB and fresh-import gates when a downstream asset/export is part of the contract. Do not imply that procedural shaders, atmospheric volumes or Cycles lighting survive a GLB export unchanged.

Establish the contract

1. Resolve the exact Blender executable and record blender --version.
2. Resolve routine visual and technical choices from context and state useful defaults. Use $blender-agent-studio:blender-art-direction-intake when an unresolved decision would cause substantial rework and context provides no reasonable default, or when the user explicitly requests a brief/concept.
3. Convert the request into a short modeling contract before editing:
   • required parts and visible relationships;
   • intended style and materials;
   • intended finish quality and whether low-poly is actually requested;
   • dimensions and an appropriate triangle range or performance target;
   • moving parts, pivots, or required contexts;
   • named stages, review evidence, and any user approval gates;
   • deliverables and evidence views.
4. Keep subjective goals as explicit review questions. Do not silently turn them into arbitrary geometry thresholds.
5. Read references/modeling-contract.md for the contract shape.

Default to a finished-quality asset

For new subjects, follow the shared guidance's reference-gathering step before the graybox. Write down the proportions and relationships the references imply, then compare matching graybox views against them. A folder of images without observations or a comparison pass does not complete the reference stage.

Treat “game-ready,” “stylized,” and “optimized” as quality constraints, not as synonyms for visibly low-poly.

• Unless the user explicitly requests low-poly, blockout-only, voxel, faceted, PS1-era, or an unusually strict platform budget, target a polished smooth model with clean silhouettes, bevels or support geometry, appropriate subdivision or curve resolution, and readable secondary and tertiary forms.
• Use the lowest density that preserves the intended finish from every evidence view. Do not optimize away the shape language, material breaks, or contact detail that makes the asset feel complete.
• Preserve an explicitly requested low-poly style. Do not smooth or subdivide away intentional planar forms merely because the normal default is polished.
• A triangle ceiling is a limit, not a target. Do not celebrate being far under budget when the result still reads as a blockout.

For characters, also use $blender-agent-studio:blender-character-workflow and its form/fit review. Low-poly style is not an exception to plausible torso and shoe proportions, connected garment surfaces, or fitted accessories. A beveled box and detailed texture do not by themselves resolve a blocky form.

Work in named stages

Read references/staged-quality-workflow.md and make the current stage explicit in progress updates and final_report.md. For a normal finished asset, satisfy all stages' exit criteria. Adjacent stages may share a build and review pass; a repair revisits only affected stages:

1. contract and references;
2. graybox and proportion;
3. primary and secondary forms;
4. structural refinement and production topology;
5. UVs, materials, and textures;
6. tertiary detail, smoothing, subdivision, and presentation polish;
7. export, fresh-import validation, and final evidence.

Do not add final materials to disguise unresolved proportions or unsupported parts. Do not call a graybox or refined blockout “finished.” If the user asks for approval-gated iteration, stop after the requested stage, show multiple angles, and wait for approval. Otherwise, including interactive work, perform self-review at each relevant milestone and continue without asking.

Author for iteration

1. Create one deterministic entry script. Set seeds explicitly when randomness is used.
2. Start from a clean scene and name semantic parts, assemblies, materials, actions, cameras, and anchors.
3. Model readable primary forms before small surface detail. For multi-zone scenes, plan finish coverage for the whole contracted asset. Review secondary work areas and inspectable reverse/top surfaces before spending the remaining detail budget on the hero object. Authored-camera beauty renders supplement the promised multiview checks.
4. Give every visibly moving or functional part a plausible connection, support, guide, hinge, sleeve, rail, or parent.
5. Keep important dimensions and animation frames as named constants near the top of the script.
6. Preserve editable construction where useful, but evaluate modifiers before measuring exported geometry.
7. Use bevel, subdivision, weighted normals, smooth shading, curve resolution, or deliberate manual topology according to the requested style. Inspect the evaluated result rather than assuming a modifier equals polish.
8. Read references/procedural-patterns.md when implementing reusable Blender helpers.

Execute and inspect

Run Blender headlessly:

$env:BLENDER_EXECUTABLE = "C:\path\to\Blender\blender.exe"
& $env:BLENDER_EXECUTABLE `
  --background --factory-startup --python .\create_asset.py

If blender is already on PATH, use it directly. The bundled MCP and benchmark runner also accept an explicit blenderPath or --blender value.

After each coherent geometry change, regenerate the authored asset and inspect the affected numerical invariants and low-cost views. At a quality milestone, open multiview evidence and assess silhouette, proportion, supports, intersections, readability, orientation, and requested details. Open individual views at original detail when the contact sheet cannot resolve a defect.

Use $blender-agent-studio:blender-asset-validation for early fresh-import checks when a change risks export behavior and for full final inspection and standardized evidence. A cosmetic iteration does not require the entire export pipeline unless it changes the exported appearance. Refine durable source and regenerate; do not patch generated outputs manually.

For animated or articulated assets, also use $blender-agent-studio:blender-animation-workflow.

Completion gate

Do not call the model complete until:

• the source script reruns from a clean Blender process;
• generated artifacts open after fresh GLB import;
• technical gates appropriate to the task pass;
• every required stage reached its exit criteria or was explicitly excluded by the user;
• the asset no longer reads as a blockout from any required view unless a blockout was the requested deliverable;
• curves and broad surfaces are smooth enough for the intended view distance, while explicitly low-poly forms retain their intentional faceting;
• materials describe the requested substances and visible UV, texture, shading, or lighting failures are resolved;
• required parts and spatial relationships are visible from the evidence views;
• no major component reads as floating, accidental, or mechanically unexplained;
• the final response includes the source, .blend, .glb, metrics, and rendered evidence paths.