Build Rigged Game Assets

Establish the contract first

• Locate the source image/model, runtime loader, skeleton conventions, action API, equipment system, collision system, catalog schema, review routes, budgets, and tests.
• Classify the actor as

  character

  or

  monster

  .
• Classify delivery as

  catalog-only

  ,

  review-only

  , or

  runtime

  . Do not call an asset implemented until gameplay loads it.
• Preserve immutable source artifacts separately from optimized runtime files.
• Decide target formats before using a credit-bearing generator.

assets/

• character-asset-manifest.template.json

• monster-asset-manifest.template.json

scripts/validate_asset_manifest.py <manifest>

--root <repo>

--require-verified

Build the common actor package

1. Preserve the source.
   • Record reference images, source tasks, imported files, licenses, generator/model, and material provenance.
   • Never overwrite the high-detail source with the runtime optimization.
2. Produce the main model.
   • Normalize scale, Y-up, forward axis, pivot, ground contact, materials, textures, and triangle budget.
   • Ground from feet or authored contact sockets, never a weapon tip, cloth tail, shadow, or VFX.
   • Remove isolated triangles, non-manifold fragments, duplicate surfaces, hidden generator debris, and unused nodes.
3. Build one stable rig.
   • Name root, hips/center mass, head, limbs, feet, and attachment sockets explicitly.
   • Keep at most four normalized skin influences per vertex unless the engine contract says otherwise.
   • Preserve a deterministic bind pose, finite transforms, positive usable scale, and compatible skeleton identity across every clip.
4. Build the action library.
   • Give every clip a stable

     id

     , semantic

     role

     , loop policy, root-motion policy, duration, contact events, and source.
   • Prefer animation-only GLBs for separate clips. They must not smuggle duplicate meshes, materials, or textures.
   • Make non-looping actions restartable and return to the intended locomotion/idle state.
5. Build runtime integration.
   • Use one asset-specific loader/factory with disposal for geometry, materials, textures, mixers, events, observers, renderers, and animation frames.
   • Publish source, model, rig, socket, action, equipment, collider, triangle, material, and status metadata.
   • Keep a truthful rollback path when replacing a shipped runtime asset.

Apply the character profile

• Declare every canonical slot: headgear, vestment, gloves, leggings, boots, main hand, offhand, and back/ranged.
• Mark unsupported slots explicitly instead of silently omitting them.
• Store supported swords, shields, bows, helmets, armor, and accessories as separate assets or independently addressable roots.
• Bind equipment to named sockets. Declare active and stowed sockets when an item changes state.
• Toggling a slot off must leave no floating trim, grip, one-triangle shards, duplicate baked weapon, or incidental gameplay prop.
• When the source contains no body beneath occluded equipment, create a fitted rig-following underlayer and record that limitation. Do not invent invisible anatomy claims.
• Require the character action roles

  idle

  ,

  walk

  ,

  run

  ,

  primary-attack

  ,

  hit

  ,

  dodge

  , and

  death

  . Add block/ward, charged attack, cast, bow aim/shoot, jump, interaction, or class actions when gameplay exposes them.
• Keep equipment visibility independent from action selection unless the action deliberately changes active/stowed state.

Apply the monster profile

• Do not require wearable slots. Separate only weapons, breakable parts, dropped parts, projectiles, or props that must animate, hide, detach, or collide independently.
• Require the monster action roles

  idle

  ,

  locomotion

  ,

  primary-attack

  ,

  hit

  , and

  death

  .
• Add secondary attack, ranged attack, pounce, cast, summon, block, stagger, roar, spawn, or phase-transition roles only when the combat design uses them.
• Define navigation and hurt colliders plus per-attack telegraph, active, and recovery windows.
• Bind attack origins, projectiles, VFX, audio, head/torso targeting, and feet to named sockets.
• Keep visual animation separate from authoritative attack volumes, damage timing, AI state, and navigation.
• Test every attack at its authored range, facing, contact window, interruption, death, and cleanup boundary.

Build the catalog lab

• Catalog card: transparent PNG, truthful label, zero WebGL canvases.
• Selected inspector/review route: exactly one live model canvas with source, status, budgets, rig, sockets, actions, and known boundaries.

• drag-to-rotate while preserving page scroll;
• action selection for every shipped clip, including visible Walk, Hit, and Death controls;
• restart and pause/play;
• equipment toggles for characters, including clean all-off and active/stowed states;
• combat-state or attack controls for monsters instead of character gear controls;
• a fallback image until the real model is ready;
• deterministic runtime telemetry defined in the catalog/testing reference.

Verify and commit

git diff --check

• verify the card has one transparent image and zero canvases;
• verify the inspector has one model host/canvas;
• click every action and confirm the selected/playing state;
• toggle every character slot individually and all equipment off;
• inspect front, back, and both sides for fragments, holes, grounding, and socket errors;
• drag the model, scroll the page, and test the supported narrow/mobile layout;
• verify fallback and model-load failure behavior;
• verify no broken media or console warning/error.