Forge Data

forge-data — searchable 3D reference

A curated, local dataset + a tiny BM25 search engine. It answers "give me a vetted starting point / cross-check" fast — polycount budgets, texel densities, format decisions, headless CLI flags, PBR presets — without spending model context on the entire corpus. It is the data layer of the Forge suite; the other forge-* skills remain the authorities that make the actual craft decisions.

Project memory: if FORGE.md exists at the project root, read it first — it carries the confirmed tool choices, poly budget, texel density, render engine, PBR workflow, and output paths for this project. forge-data answers "what is the standard starting point?"; FORGE.md answers "what did we already decide for this project?"

Forge suite map

• forge — router / orchestrator (entry point for all 3D work)
• forge-brief — writes FORGE.md; extracts ATELIER.md brief
• forge-standards — authoritative 3D design-token rules (units, naming, pivots, budgets)
• forge-model — bpy / bmesh polygonal modelling
• forge-parametric — OpenSCAD / CadQuery / FreeCAD code-CAD
• forge-procedural — Geometry Nodes, SDF, L-systems, scatter
• forge-topology — retopo, decimation, LOD chains, boolean cleanup
• forge-uv — UV unwrap, seams, packing, texel density, UDIMs
• forge-material — PBR shading, Principled→glTF mapping
• forge-texture — baking (normal / AO / curvature / displacement) + procedural
• forge-light — lighting rigs, HDRI / IBL, color management (AgX / ACES)
• forge-render — headless Cycles renders, turntable + contact-sheet QA
• forge-rig — armatures, IK / FK, skinning / weights, blend shapes
• forge-animate — keyframes, F-curves, baking, skeletal + morph export
• forge-sim — cloth, rigid / soft / particles, hair / fur, fluids
• forge-export — format matrix, glTF / GLB / USD / FBX export, engine conventions
• forge-optimize — gltf-transform, Draco / Meshopt / KTX2, web delivery, atelier-webgl handoff
• forge-intake — photogrammetry, Gaussian splat / NeRF, AI text-to-3D cleanup
• forge-validate — manifold, normals, scale, polycount, UV overlap, glTF-Validator, render QA
• forge-data ← you are here (BM25 reference library)

Atelier seam: forge-export → forge-optimize → Skill(atelier-webgl)

When to query (and when not)

Use for: fast budget lookup (polycount / texel density for a given platform and asset class); format selection (should this be FBX, GLB, or USD?); headless CLI flags for Blender / assimp / gltf-transform; PBR channel-pack conventions (ORM layout, color-space rules); Windows gotcha cross-checks before running a conversion.

Do not use for: making the final craft decision (use the responsible forge-* skill); reading or auditing existing files (use Read tool or forge-validate); real-time rendering parameters already locked in FORGE.md (read FORGE.md directly).

How to query (Windows: use python, not python3)

cd $CLAUDE_CONFIG_DIR/skills/forge-data/scripts
python search.py "<query>" --domain <domain> [-n 3]
python search.py "<query>"              # auto-detect domain
python search.py "<query>" --json       # machine-parseable output

Keep -n small (≤ 5). The --json flag bypasses per-field 300-char truncation — use only when the caller needs structured data. Python is stdlib-only; no network calls.

Datasets → which Forge skill owns the decision

The flow

1. Receive query from a sibling skill (e.g. forge-standards asks for mobile hero polycount, forge-export asks for FBX vs GLB for a Unity target).
2. Run scripts/search.py with the relevant --domain and the query string. Keep -n ≤ 5.
3. Read the structured output — check the Decision owner column; the result is a starting point, not a binding rule.
4. Cross-check FORGE.md — if the project has already locked a budget or format, that takes precedence over the dataset default.
5. Hand back the answer to the requesting skill. forge-data never makes the final decision itself.

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

Deep references (load on demand)

Full copy-paste-ready detail lives in references/ — the SKILL.md body only carries pointers. Load a reference file with Read when the summary row from a CSV search is not enough.

Operating principles

• Data layer, not decision layer. forge-data surfaces starting points; forge-standards, forge-export, forge-material, and the other craft skills make binding decisions.
• FORGE.md wins. Any value already locked in the project memory file overrides a dataset default. Always check FORGE.md before citing a search result as the answer.
• Stdlib only, no network. Every script is pure Python stdlib. Never make a network call from a search or probe script.
• Keep tokens small. Keep -n ≤ 5 and avoid --json unless the caller needs structured data — search results enter the model's context and large payloads waste budget.
• 2-char tokens match. The BM25 tokenizer keeps len(w) >= 2, so queries like "uv", "3d", "ao", "gi", "lm" all resolve correctly.