rig-it

rig-it

Rigging fails in chat because the model writes one script, gets an armature that binds, and never sees that the elbow is wrong. This skill never lets you judge a rig from the code that built it. Every rig and every clip goes through a contract, a family route, verified weights, numeric gates on a fresh reimport, and rendered frames you look at before anyone calls it done.

Read RULES.md before touching a mesh. Run WORKFLOW.md commands as written. Family bone lists live in references/body-plans.md.

The loop

contract  ->  route  ->  rig + weights  ->  clips  ->  export  ->  fresh reimport
   ^                                                                    |
   |                        polish passes (never thresholds)   <--  gates + frames

1. Contract first. Write rig-contract.json (schema in templates/): family, engine rig type, required clips, rigid parts with sockets, gameplay move speed. Tooling reads it; tooling never redefines it.
2. Route by body plan, not by wish.

   Body	Route	Scripts
   Biped that a Mixamo-template rigger accepts (Meshy, mixamo.com)	rig input, provider rig, preset motions, package onto the production mesh	rig_input.py, package_biped.py
   Quadruped, hexapod, octopod, flyer, maw, winged hybrid, humanoid the provider mangles	deterministic scaffold rig from bounds, bind, author contact-driven clips	scaffold_rig.py, author_clips.py

3. Weights are verified, not assumed. Automatic weights are a starting point. Unweighted vertices fall back to the nearest bone. Max 4 influences, normalized. Sole patches harden to the foot bone. Provider weights transfer onto the production mesh with both meshes in rest pose.
4. Stationary navigation root. Root carries no skin weight and never moves. Body motion lives on the body bone. Root motion is off in engine.
5. Rigid parts are skinned 100% to one hand bone and seated with a grip rule. Never bone-parented (FBX pivots differ).
6. Export, then judge the reimport. export_fbx.py writes one FBX per clip plus the production multi-take FBX. review_clips.py reimports each clip into an empty scene, measures every frame on real vertices, renders frames over a checker floor with the camera following travel, and runs gates.py.
7. Fix in the package, never in the thresholds. Provider walks slide, deaths sink, idles hitch. package_biped.py polish passes fix those in order. Gate thresholds change only with a recorded decision.
8. Reel to the reviewer. reel.py cuts labeled 3 s sections. A passing gate is screening, not art acceptance. The human signs off, the model does not.

Quick start

Iterate as --version 2, 3. The pipeline refuses to overwrite a version; earlier versions are the trail.

Per-subject checklist

1. Contract written and python3 scripts/contract.py rig-contract.json prints the required bones.
2. inspect_model.py shows zero unweighted vertices, every required bone, every rigid part, +Z forward, feet on the ground plane.
3. Clips authored or packaged; receipt has contacts and nominal speed for Locomotion.
4. review_clips.py gates pass. Look at the frames from front and side, not only the review camera. Held parts hide behind bodies; legs pulled under hips read as glued.
5. Reel sent to the reviewer. Record the version, the passes used, and the findings.
6. Engine install per references/unity-import.md: Humanoid avatar validates or Generic import binds by path, Apply Root Motion off, Locomotion rate bound to moveSpeed / nominalSpeed.

Do not

• Judge a rig from the scene that authored it. Export, reimport, then judge.
• Raise a gate threshold to pass. Fix the package.
• Bone-parent a weapon. Skin it to one bone.
• Transfer weights with either mesh posed.
• Fake gameplay speed with a global animator speed.
• Claim art acceptance from a technical pass.
• Upload assets you do not own to a rigging provider, or publish them there.