# 3d Asset Production > Use this skill to turn generated, captured, scanned, or modeled 3D output into production-ready standalone assets for DCC, real-time engine, web, or interchange delivery. It covers requirements, topology repair, scale and axes, UVs, PBR baking, LODs, static and skeletal readiness, rig handoff, optimization, glTF/USD/FBX delivery, validation, QA, and rights/provenance. Do not use it for provider-s… - **Type:** Skill - **Install:** `agentstack add skill-calesthio-generative-media-skills-3d-asset-production` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [calesthio](https://agentstack.voostack.com/s/calesthio) - **Installs:** 0 - **Category:** [Developer Tools](https://agentstack.voostack.com/c/developer-tools) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [calesthio](https://github.com/calesthio) - **Source:** https://github.com/calesthio/generative-media-skills/tree/main/skills/production/3d-craft/3d-asset-production - **Website:** https://github.com/calesthio/OpenMontage ## Install ```sh agentstack add skill-calesthio-generative-media-skills-3d-asset-production ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # 3D Asset Production This skill is for finishing an asset so another artist, engine, viewer, marketplace, configurator, game, AR app, or renderer can actually use it. The input may be generated mesh output, photogrammetry, NeRF/GSplat-derived mesh, scan cleanup, sculpt, CAD-derived polygon export, hand-modeled geometry, or a kitbashed draft. The output is a documented, validated asset package with predictable geometry, scale, materials, texture maps, LODs when needed, and provenance. Use it when the user asks for asset prep, retopology, cleanup, optimization, baking, UVs, PBR textures, LODs, rig handoff, static mesh delivery, skeletal mesh readiness, game-ready assets, web-ready assets, GLB/glTF/USD/FBX export, DCC-to-engine delivery, or QA of a 3D model. Do not use it for: - Choosing or prompting a specific 3D generation provider. - Building a whole environment, shot, level, or composed scene. - Running motion-capture sessions or solving mocap performances. - Exact CAD, toleranced manufacturing, BIM compliance, or simulation-grade engineering geometry. - Character performance direction, although this skill can prepare a mesh for rigging or animation handoff. ## Start With The Delivery Contract Before touching topology, define the asset's consumer. A beautiful mesh is still a failed delivery if the receiving engine cannot load its textures, its unit scale is wrong, or the rigging team cannot bind it. Ask for or infer these requirements: - Asset role: hero prop, background prop, avatar body, modular garment, weapon, vehicle, scan archive, AR product, web configurator item, cinematic asset, marketplace asset, mobile asset. - Target consumers: Blender, Maya, Houdini, Unreal, Unity, WebGL/three.js/Babylon, USD pipeline, DCC handoff, custom runtime, marketplace. - Format set: source DCC file, `.glb`/`.gltf`, `.usd`/`.usda`/`.usdc`/`.usdz`, `.fbx`, texture archive, preview renders, validation report. - Runtime budget: triangle count, vertex count, material count, draw-call/submesh target, texture resolution, compressed texture formats, LOD count, memory/download size, supported extensions. - Spatial contract: real-world dimensions, unit system, axis/up convention, pivot, origin, bounds, snap points, socket points. - Material contract: PBR workflow, color space, texture channel packing, transparency mode, normal tangent basis, renderer-specific shader expectations. - Animation contract: static mesh, skeletal mesh, blend shapes/morph targets, cloth/hair handoff, deformation range, max bone influences, root bone, naming convention. - Rights contract: input source, user rights, third-party references, scan consent, generated-output terms, trademarks/logos, attribution, license, distribution region. Treat unknown target requirements as a risk, not as permission to guess. If the user cannot specify budgets, create a defensible baseline and label it as a proposed budget. ## Evidence Classes This skill uses three kinds of claims: - Documented facts are statements from official specifications, vendor manuals, standards, or project documentation. They are cited in the source notes. - Empirical observations are repeatable production findings that should be verified on the user's actual asset and target runtime. - Production heuristics are practical defaults. They are not universal rules; change them when the delivery contract says otherwise. ## Documented Facts To Anchor Decisions - glTF 2.0 is an API-neutral runtime asset delivery format, not an authoring format. It represents scene descriptions with JSON plus binary buffers and image resources, or as GLB for a single binary container. Source: Khronos glTF 2.0 specification, verified 2026-07-11, https://registry.khronos.org/glTF/specs/2.0/glTF-2.0.html. - glTF uses a right-handed coordinate system with +Y up, +Z forward, -X right, and meters for linear distances. Source: Khronos glTF 2.0 specification, verified 2026-07-11. - glTF material core is metallic-roughness PBR. Base color textures are sRGB; metallic-roughness texture data is linear, with roughness in green and metalness in blue. Normal textures are tangent-space and linear. Source: Khronos glTF 2.0 specification, verified 2026-07-11. - glTF skinning uses linear blend skinning with `JOINTS_n` and `WEIGHTS_n` attributes; one joint/weight set gives up to four influences per vertex, and `JOINTS_n` and `WEIGHTS_n` sets must match. Source: Khronos glTF 2.0 specification, verified 2026-07-11. - glTF node transforms can use either a matrix or TRS. Animated nodes must not use `matrix`; animations target translation, rotation, scale, or morph target weights. Source: Khronos glTF 2.0 specification, verified 2026-07-11. - glTF Validator validates JSON/GLB correctness, internal references, binary buffers, forbidden accessor values such as NaN, accessor bounds, sparse accessors, animation data, images, and several extensions. Source: KhronosGroup/glTF-Validator README, verified 2026-07-11, https://github.com/KhronosGroup/glTF-Validator. - USD is designed for scalable interchange, composition, layering, overriding, and organization of 3D scene description across DCC pipelines. USD can represent models, geometry, shading, animation, primvars, payloads, references, variants, and asset metadata. It is not itself a high-performance rigging system. Source: OpenUSD Introduction, verified 2026-07-11, https://openusd.org/release/intro.html. - Unreal Engine's FBX pipeline imports static meshes, skeletal meshes, animation, morph targets, materials, textures, and multiple LODs, and Unreal documents that its FBX import pipeline uses FBX 2020.2; exporting with a different FBX version may cause incompatibilities. Source: Unreal Engine FBX Content Pipeline, verified 2026-07-11, https://dev.epicgames.com/documentation/en-us/unreal-engine/fbx-content-pipeline. - Unreal skeletal mesh assets contain visual mesh geometry and skeleton bone data; they are usually authored in external DCC software and imported as FBX. Source: Unreal Engine Skeletal Meshes documentation, verified 2026-07-11, https://dev.epicgames.com/documentation/en-us/unreal-engine/skeletal-mesh-assets-in-unreal-engine. - Unity model files can contain meshes, animation rigs/clips, materials, and textures. Unity's primary model-file support is FBX; Unity recommends direct FBX export from the DCC into the project in most cases. Source: Unity Manual, Importing a model, verified 2026-07-11, https://docs.unity3d.com/Manual/ImportingModelFiles.html. - Unity rig import settings distinguish None, Legacy, Generic, and Humanoid animation types. For generic rigs, Unity needs a root node; skin weights affect per-vertex bone influence count and performance. Source: Unity Manual, Rig tab Import Settings reference, verified 2026-07-11, https://docs.unity3d.com/Manual/FBXImporter-Rig.html. - meshoptimizer provides mesh indexing, vertex cache optimization, overdraw optimization, vertex fetch optimization, quantization, simplification, compression, analyzers, and glTF-related tooling. Its recommended core pipeline orders indexing before cache optimization, optional overdraw optimization, fetch optimization, quantization, filtering, and optional shadow indexing. Source: meshoptimizer README, verified 2026-07-11, https://github.com/zeux/meshoptimizer. ## Production Heuristics That Usually Hold - Finish the asset against the weakest target, not the strongest DCC. If a mobile WebGL viewer and Unreal are both targets, the web budget usually constrains geometry, material count, extensions, and texture sizes. - Preserve an editable source file and export a runtime derivative. Do not make GLB/FBX/USDZ the only source of truth unless the user explicitly wants a flat interchange package. - Separate semantic meshes only when downstream operations need it: material assignment, sockets, destructibility, animation, collision, occlusion, user customization, or culling. Splitting everything by generated fragments usually damages runtime performance and asset usability. - Prefer indexed triangle meshes for real-time delivery unless the target specifically requires another topology. Keep curves, subdivision, hair, procedural modifiers, and high-poly sculpt data in the source package, not the runtime export, unless the consumer supports them. - Bake detail from high-poly/generated/scanned source to a clean low/mid-poly render mesh. Runtime assets should usually carry high-frequency detail in normal, ambient occlusion, curvature/cavity, and material maps rather than in noisy dense topology. - Protect silhouette, deformation zones, contact surfaces, sockets, and UV/material boundaries during simplification. Let interior flat regions and visually hidden backs reduce first. - For deformation, regular edge flow and stable weights matter more than raw triangle count. A slightly heavier mesh with clean loops can animate better than an aggressively decimated mesh. - For hero assets, verify by import into the actual target engine/viewer. Validator success is necessary but not enough for visual parity. ## Intake Triage Classify the source before deciding the repair path. ### Generated Mesh Output Common traits: uneven tessellation, fused parts, non-manifold pockets, noisy normals, arbitrary scale, missing UV intent, baked colors instead of material logic, plausible silhouette with weak mechanical construction. Best path: 1. Freeze a source archive and record generation provider/model/settings if known. 2. Decide whether the mesh is a concept reference, a high-poly bake source, or a direct cleanup candidate. 3. If topology is chaotic but shape is valuable, retopologize rather than decimate-only. 4. Rebuild UVs and PBR maps from the approved surface appearance. 5. Validate by import, not by visual inspection in the generator's viewer. ### Captured Or Scanned Output Common traits: high density, holes, scan shadows, texture-lighting baked into albedo, fragile thin surfaces, scale drift, scan table/floor debris. Best path: 1. Preserve raw scan/photogrammetry source. 2. Establish real scale from a measured feature or calibration object. 3. Remove debris and reconstruct missing geometry only where physically justified. 4. Build clean retopology or optimized decimation depending on role. 5. De-light textures if the asset must respond to new lighting. ### Modeled Or Sculpted Output Common traits: cleaner intent, possible unapplied transforms/modifiers, hidden high-poly details, overlapping UV shells, unsupported procedural materials. Best path: 1. Apply/export only intentional transforms and modifiers. 2. Keep authoring file with non-destructive layers where useful. 3. Bake high-poly or procedural detail to maps for runtime targets. 4. Validate normals, tangents, UV sets, pivots, and unit scale after export. ### CAD-Derived Polygon Output Common traits: exact design intent but unsuitable tessellation, many parts/material IDs, tiny bevels, n-gon/trim conversion artifacts, real-world units. Best path: 1. Do not claim manufacturing exactness after polygon conversion. 2. Retain CAD as provenance/reference if licensed. 3. Tessellate with chord/angle tolerances appropriate to the viewing distance. 4. Merge or instance repeated parts where runtime permits. 5. Rebuild materials as renderable PBR approximations. ## Geometry And Topology Repair The goal is not a theoretically perfect mesh; it is a mesh whose topology supports the intended rendering, deformation, collision, LOD, and editing tasks. Minimum repair checklist: - Remove duplicate vertices where they are not required for UV, material, hard-normal, tangent, color, or skinning boundaries. - Delete hidden debris, isolated fragments, accidental internal shells, and zero-area or near-zero-area triangles. - Resolve non-manifold edges, self-intersections, flipped normals, open holes, and inconsistent winding according to asset intent. - Preserve hard-surface support loops or bevels that define silhouette and specular response. - Avoid long skinny triangles in deformation areas and high-specular surfaces. - Keep material boundaries and UV seams intentional, named, and documented. - Use normals deliberately: flat for planar mechanical surfaces, smoothed for continuous curved surfaces, custom/weighted normals when they are part of the shading design. Retopology decision: - Retopologize when the asset must deform, receive clean bakes, support predictable LODs, use mirrored/symmetrical UVs, or survive close camera inspection. - Decimate when the asset is static, scanned, mostly organic, non-deforming, and the source detail is more important than editability. - Re-model when the source is visually plausible but physically incoherent, such as generated machinery with fused impossible parts. Topology QA: - Inspect wireframe over shaded view from close and far distances. - Check silhouette at target FOV and distance. - Test smoothing under high-contrast HDRI or studio lights. - Confirm no surface tears after export/import round trip. - For deforming assets, test extreme bend/twist poses before signoff. ## Scale, Axes, Origin, Pivots, And Naming Scale errors are expensive because they propagate into physics, animation, camera framing, lighting, AR placement, and UI fit. Set and document: - Unit basis: meters, centimeters, or target-engine units. - Real dimensions: bounding box and key measured dimensions. - Orientation: up axis, forward axis, front-facing convention, and any format conversion. - Pivot/origin: bottom center for props that sit on surfaces, center of mass for physics props, hinge point for doors/wheels, grip point for weapons/tools, root at floor between feet for characters unless target says otherwise. - Transforms: freeze/apply scale and rotation before export unless the target pipeline needs authored transforms. - Naming: stable ASCII names for objects, materials, bones, sockets, UV sets, textures, and LODs. Avoid autogenerated gibberish and duplicate material names. For glTF delivery, remember the documented coordinate and unit contract: right-handed, +Y up, +Z forward, meters. If the authoring DCC or engine uses a different convention, explicitly test the exporter/importer conversion. ## UV Strategy UVs are a delivery system for texture resolution, seams, baking, and material reuse. Choose UV layout by asset role: - Unique hero asset: non-overlapping UV0 for baked PBR maps, consistent texel density, shells aligned for paintability, seams hidden on backs/creases/material breaks. - Tiled material asset: UV0 can use repeatable coordinates; create a second non-overlapping UV set for lightmaps or baked masks if required by the engine. - Mirrored character or prop: mirroring is acceptable for symmetric albedo/roughness but risky for readable logos, asymmetrical damage, tangent-space normal maps, and baked AO. Split or offset mirrored areas as needed. - Modular kit part: snap-friendly UVs, consistent texel density across modules, trim sheets when the project already uses them. - Scan-derived asset: unwrap for stable reprojection; clean baked lighting or shadows from albedo if relighting is required. UV QA: - No unintended overlaps in a bake UV set. - Adequate padding for the final mip chain and texture resolution. - Consistent texel density for surfaces viewed together. - Seams placed where geometry, material, or visibility already hides them. - Checker texture shows no severe stretching, flipped islands, or mis-scaled shells. - Target engine recognizes every UV set required by materials, lightmaps, decals, or runtime effects. ## PBR Materials, Baking, And Texture Packaging Build material output from the target shading model, not from whatever the source happened to contain. Common PBR map set: - Base color/albedo: surface color only, no baked direct lighting unless the target intentionally wants a lit texture. - Normal: tangent-space for real-time assets unless the target requests object/world-space normals. - Roughness … ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [calesthio](https://github.com/calesthio) - **Source:** [calesthio/generative-media-skills](https://github.com/calesthio/generative-media-skills) - **License:** MIT - **Homepage:** https://github.com/calesthio/OpenMontage Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-calesthio-generative-media-skills-3d-asset-production - Seller: https://agentstack.voostack.com/s/calesthio - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.