Forge Standards

Forge Standards — 3D Design Tokens

The ground-truth layer. Every other Forge skill ingests these constants before writing a line of code. Without them, two independently authored assets collide in the same scene — wrong scale, mismatched texel density, broken LOD slots — and the render-verify step cannot produce a stable comparison baseline.

Project memory: if FORGE.md exists at the project root, read it first — it carries the confirmed coordinate system, scale unit, poly budget, texel density, render engine, and output paths for this project. If FORGE.md is absent, this skill writes the standards sections of it. When ATELIER.md exists, check for a ## Forge 3D assets block.

Suite map — key cross-references:

• forge (router) dispatches here after forge-brief; reads FORGE.md target engine.
• forge-brief writes FORGE.md; calls this skill to populate budgets + coordinate system.
• forge-model / forge-parametric / forge-procedural consume naming + pivot rules.
• forge-uv consumes texel density targets, UV utilization floors, and channel-pack policy.
• forge-texture consumes ORM packing policy and texture suffix table.
• forge-topology / forge-export consume handedness rules, axis-swap matrices, and LOD ratios.
• forge-validate is the enforcement gate — it runs checks against every rule defined here.
• atelier-webgl receives GLB assets; its Three.js scenes use glTF Y-up RH, matching this skill's web profile.
• atelier-perf-a11y is the web-runtime performance gate after forge-export/optimize.

Run = call the Skill tool with the exact name. Writing "now run forge-validate" in prose runs nothing.

Decide first: which target profile?

Before consulting any table below, confirm:

1. Read FORGE.md (if absent, invoke Skill("forge-brief") to write it first): extract ## Target, ## Coordinate system, ## Budgets.
2. Pick the profile that governs this asset:

Blender always authors in Z-up, right-handed, meters. Every axis-swap and unit correction is baked at export time, not during modeling. Full axis-swap matrices and conversion snippets: references/coordinate-systems.md.

The flow

1. Read FORGE.md — load existing profile; if absent, invoke Skill("forge-brief") to create it first (do not just narrate the handoff).
2. Confirm target profile (table above) — this gates every budget and convention downstream.
3. Generate FORGE_STANDARDS.json at project root — the machine-readable token file. Schema and full example: references/forge-standards-schema.md.
4. Apply naming convention — use the prefix + suffix tables; verify with the regex in the schema. Full prefix/suffix tables: references/naming-conventions.md.
5. Set pivot / origin per object class — ground props go bottom-center; hinged objects go at the hinge. Pivot rules + bpy snippets: references/pivot-origins.md.
6. Load polycount budgets for the confirmed profile — PC/console, mobile, web, or Nanite-era. Budget tables + LOD ratios: references/polycount-budgets.md.
7. Load texel density targets — px/m tier per asset class, UV utilization floor, channel-pack policy. Density tables + calc snippet: references/texel-density-uv.md.
8. Validate existing assets — run forge-validate (call via Skill tool) against the written standards. Headless validation invocation: references/validation-headless.md.
9. Render-verify scale — after export, render against a 1.8 m human silhouette at world origin with fixed, reproducible settings: Cycles 16 samples, seed=0 (or Workbench), 512×512, fixed orthographic side camera. Read the PNG and compare silhouette height. Detect→fix: silhouette ≈ half (or 100×) expected height → unit mismatch → set unit_settings.scale_length + transform_apply(scale=True); asset rotated 90° → Z-up vs Y-up export → re-export with export_yup=True. Settings + table: references/validation-headless.md (deeper render-QA: forge-validate render-qa-guide).

Coordinate systems & axis-swap in brief

Model in Blender Z-up RH, meters always. Export corrections per target:

The Blender glTF exporter and FBX exporter handle these automatically when given correct axis settings. For skeletal meshes to UE5: set bpy.context.scene.unit_settings.scale_length = 0.01 before export to avoid the 100× skeleton bug (static mesh imports fine, skeleton comes in at 1% scale). Full matrices, Python conversion functions, and gotchas: references/coordinate-systems.md.

Naming — quick reference

Pattern: [TypePrefix]_[AssetName]_[Descriptor]_[VariantNumber] — no spaces, no Unicode, PascalCase asset name, zero-padded number (_01).

Key prefixes: SM_ static mesh · SK_ skeletal/skinned · T_ texture · M_ material · MI_ material instance. Texture suffixes: _BC base color (sRGB) · _N normal (Linear) · _ORM packed AO/Rough/Metal (Linear). Collision (UE5 auto-detect): UCX_SM_Barrel_Oak_01 (convex hull) · UBX_ (box) · USP_ (sphere). LOD slots: SM_Barrel_Oak_LOD0 → SM_Barrel_Oak_LOD3.

Full prefix table, texture suffix table, Blender Studio in-file prefixes, and regex patterns: references/naming-conventions.md.

Polycount & LOD — quick reference

All counts in triangles (GPU draw unit). Quads ≈ half the tri count.

Default LOD ratios: [1.0, 0.5, 0.2, 0.05] (LOD0 → LOD3). Always decimate from LOD0, not cascading from the previous LOD. Full per-class tables and Nanite notes: references/polycount-budgets.md.

Texel density — quick reference

Formula: TD (px/m) = texture_resolution_px / object_dimension_m

ORM channel-pack: R = AO, G = Roughness (most bit precision under BC compression), B = Metallic. One ORM replaces three maps, saving ~4 MB per 2048² material at BC1. All tiers, UV utilization floors (≥ 85% hero, ≥ 75% env), lightmap UV rules, and bpy snippets: references/texel-density-uv.md.

Operating principles

• One profile, one truth. Every project has exactly one FORGE_STANDARDS.json at root. All Forge skills read it; none hardcode budgets or names. If the file is absent, this skill writes it first.
• Model in meters, convert at export. Blender stays Z-up RH meters throughout. Axis swaps and unit corrections happen in the exporter, never during modeling — so the source .blend is always portable.
• Apply transforms before any export. Every time. Unapplied scale is the #1 pipeline killer: it corrupts physics, normals, and skeletal animation. Automate apply_transforms() in the pre-export script.
• Naming is a contract. SM_, SK_, T_ prefixes and _BC, _N, _ORM suffixes are not style — they drive engine auto-import, LOD slot detection, lightmap UV assignment, and forge-validate regex. A single misnamed asset silently breaks the import pipeline.
• Validate before handing off. Run Skill("forge-validate") against every asset before forge-export. The standards file is machine-readable precisely so that the validation gate can run without human review.