Custom 3D RTS Asset Pipeline: A Reusable Skill Pack

---
name: custom-rts-asset-pipeline
description: Coordinate a reusable custom-art 3D RTS asset pipeline from existing models through rigging, animation, optimization, and runtime verification. Use when building an asset workflow or a repeatable game-production pack.
---

# Custom RTS asset pipeline pack

Preserve the owner's chosen artwork and reference style. Existing animated packs may supply motion or rig references; substituting their visible characters does not prove custom-character readiness.

Before creating infrastructure, inventory installed skills, runner scripts, pinned images, saved job configurations, retained GPU volumes, model caches, and completed derivatives. Verify availability instead of assuming a stopped GPU can resume. Reuse tested operations; add only the missing stage. Do not rent multiple GPUs to solve unproven mesh or animation problems.

## Stage contracts

| Stage | Input → output | Completion evidence |
|---|---|---|
| Reference and concepts | reference video/art → approved character/building concepts | silhouettes and style match at intended camera distance |
| Mesh generation | approved art → textured static mesh | intact limbs and materials, multiview review |
| Readiness | existing mesh → inspected derivative and issue ledger | topology, pose, triangle count, pivot, axes, missing skins/actions recorded |
| Rigging | mesh → skeleton → weighted/skinned mesh | real joints, skin weights, and visible deformation; appearance retained |
| Motion | rig + action source → baked clips | idle/walk/attack/hit/death; role-specific work/heal/ranged actions where needed |
| Export | animated mesh → optimized GLB | glTF validation; skins, joints, clips, sockets and materials survive |
| Runtime | GLB → browser game unit | actual actions, feet/pivot, team colors, attachments, crowd performance |

Prove one existing custom character across the stages before batch conversion. Record failures and repairs alongside successes. Separate a skeleton-only asset, a skinned asset, a motion test, and an accepted game character.

## Available focused guides

- UniRig: https://github.com/VAST-AI-Research/UniRig — skeleton and skin-weight inference; motion still requires a separate step. Companion skill: [UniRig character pipeline](https://promoteindex.com/skills/unirig-rig-custom-3d-characters-without-replacin).
- Existing-mesh readiness: companion skill `rts-asset-readiness`, including source/license evidence and GLB metadata auditing.
- Existing image-to-3D generation: https://promoteindex.com/skills/run-comfyui-workflows-on-runpod-gpus-pay-per-min — reuse when generation is actually required; it does not produce character rigs.
- Export optimization: https://gltf-transform.dev/cli and https://github.com/KhronosGroup/glTF-Validator . Decode compression before feeding unsupported importers; restore runtime compression only after proving rigs and clips survive.

## Evidence and stopping rules

Each run records input hashes, repository/checkpoint revisions, installed dependencies, exact commands, output hashes/counts, visual checks, elapsed time, cost and compute stopped state. Credentials and private connection state stay outside public skill content.

A playable demo with stock units is game-logic evidence; custom-character deformation and motion require their own acceptance evidence. Publish only tested claims. For unfinished steps, state the exact missing test. Build/console checks do not establish visual fidelity or animation quality.

Initial pack assembled 2026-10-19. Eight custom humanoid web exports now retain skins and retargeted clips; the active settlement worker uses a 12-clip custom villager with separate tool and cargo props. Browser checks prove finite deformation, actual task/prop switching, scale, grounding, death handling and stock fallback behavior in the earlier combat demo. A local settlement prototype also exercises resource gathering/delivery, travel/construction, timed villager production, fog and placement validation. Original resource/camp art, polished role animation, cavalry and crowd performance remain unfinished. This is a tested baseline workflow, not a claim of production-quality game art or a complete RTS.


## Runtime lessons exercised in the settlement baseline

- Separate asset readiness from gameplay readiness. Rigged villagers do not implement gathering; a gather loop requires finite deposits, worker tasks, carrying, drop-off stock changes and real input tests. Record these as separate acceptance gates.
- Preserve a single source manifest with file hash, skin/joint count, clip names/durations, source motion/license, forward axis, sockets, units and defects. Re-exported material/socket changes alter hashes even when geometry looks similar; wait for a final manifest before copying the complete runtime set.
- Forward vectors from Blender Z-up must pass through the glTF Y-up conversion before runtime yaw correction. In the exercised export, (x,y,z) becomes (x,z,-y); align this converted forward to runtime +Z. Ground/center from actual rest bounds, then normalize the character to its metre-height contract.
- Tools and carried loads are separate low-poly GLBs. Attach them to explicit hand/chest sockets on each cloned skeleton, select visibility from the actual simulation task and resource cargo, and verify one visible prop per action with none during death. Test independent instances so props never accumulate or attach to another unit's skeleton.
- Metre-authored props need world-scale verification after character normalization. A properly authored socket can make local identity correct; otherwise compensate the actual parent world scale. Do not apply both a nominal scale correction and another unconditional multiplier. Inspect gripping and cargo at several action times; an attachment can be structurally valid but float near the face.
- Audit occlusion under runtime lighting. Baked dark color plus another occlusion texture can make custom units nearly black. Review an authored derivative with corrected AO and compare against the original concept; preserve the original source. A validator does not detect this visual defect.
- Retargeted combat swings can supply rough chop/build/mine motions, but label them as reused baseline motion. Open hands, fused props, skirt stretching and foot slide remain acceptance issues even when animation data is valid.
- Verify real pointer actions against the combined game, not only isolated model viewers. Keep a recorded settlement loop and production-bundle asset check. If waiting is accelerated through the production simulation, state that clearly; do not present it as real-time pacing or a hardware benchmark.
- Edge scrolling must track the pointer across HUD overlays. Software-rendered browser tests should wait for actual frame/camera progress rather than infer failure from a short fixed wall-time sleep.
- GPU restart and timed-stop guards need a populated pod ID, provider start time and current-run deadline. A stale STOPPED marker or missing creation timestamp invalidates the guard. Verify the live guard process and provider stopped receipt independently. No available capacity means no completed conversion; keep temporary map blockouts explicit until authored meshes arrive.