Reviewing an AI-Generated Creature Skeleton Before Rigging It

Don't send a generated creature mesh straight to rigging. First queue a four-view anatomy inspection, read back the annotated views and the editable joint profile with its uncertainty markers, correct anything wrong in the coordinate frame the analysis returns, and only then request a rig draft with reviewed_anatomy=true. The rig draft is a separate, private artifact — a skeleton, smoothed weights, starting clips, a GLB, and editable Blender source — and it still needs its own deformation and animation review before anyone calls it game-ready. Nothing in this pipeline poses fingers, builds a face rig, or ships finished combat animation.

This is the core of a non-humanoid creature rigging workflow: you're not correcting a broken biped template, you're verifying that the tool understood how many limb chains this specific body has, where the joints actually sit, and whether the mesh is grounded enough for a leg chain to get walk/run cycles at all.

Step 1 — Queue the four-view inspection

Call gripforge_creature_analyze with the source_id of the owned Library GLB. This queues a job; it does not return anatomy synchronously.

Acceptance criteria: you get back a job reference, not an inline result. If the call errors instead of queuing, confirm the source_id points to a mesh you actually own in the library — the analyzer only works on assets already ingested there.

Step 2 — Read the annotated result

Poll with gripforge_generation_read using include_preview=true. This returns:

• Annotated four-view images (front, right, back and top views) marking detected joints and limb chains
• An editable joint profile (positions, parent/child relationships, chain roles)
• Per-joint or per-chain uncertainty markers

Acceptance criteria: every limb the mesh visually has is represented by a chain in the profile, and every uncertainty flag has a plausible visual explanation (occluded limb, ambiguous tail/leg boundary, asymmetric pose in the source mesh). If a limb chain is missing, use the editable profile and current schema to correct it when the geometry is clear. If the mesh or captures are ambiguous, improve the source or rerun analysis before approving the anatomy.

Step 3 — Classify body plan and limb chains

Before judging individual joints, decide what kind of body you're looking at:

• Grounded legs — chains that touch a walking surface and are expected to carry weight (these are the ones eligible for walk/run clips with IK).
• Manipulator or non-grounded limbs — arms, wings, tentacles, tails that don't bear locomotion weight.
• Spine/torso segments — for serpentine or multi-segment bodies, confirm the segment count matches what you see, not a fixed humanoid assumption.

Example: a six-legged insectoid should show six grounded chains, not four legs plus two limbs misclassified as arms. A wyvern with two legs and two wings should show exactly that split — if a wing root is mis-tagged as a grounded leg chain, walk/run generation later will be nonsensical, so fix the classification now, not after rigging.

Acceptance criteria: chain count and role match the mesh; weight-bearing legs are identified correctly, and appendages such as a tail touching the floor are not automatically treated as locomotion legs.

Step 4 — Inspect joints and uncertainty markers directly

For each chain, check:

• Joint spacing follows visible bend points in the mesh (elbow, knee, wing-fold, tail segment boundary)
• No duplicate or collapsed joints at a single visual bend
• Uncertainty markers cluster where you'd expect (thin tails, feathered wing tips, fur-obscured joints) rather than on obviously simple limbs

If uncertainty is high on a structurally simple limb (a plain straight leg, say), inspect the mesh, captures and proposed joint locations closely. Correct a clearly identifiable mistake in the profile; rerun analysis when the evidence is insufficient.

Step 5 — Correct the profile in the returned coordinate frame

When you edit joint positions, make the edits in the same coordinate frame the analysis reported them in — don't re-derive your own world-space guess from the raw GLB. The returned frame is what the rigging step expects to consume back, and mixing frames can place joints away from their intended locations and produce incorrect deformation.

Acceptance criteria: corrected joint values still validate against the profile schema shape described in gripforge_creature_rig_schema, and re-rendering the annotated views (if supported by your tool) shows joints landing on the visible bend points.

Step 6 — Request the rig draft

Call gripforge_creature_rig with the analysis_job reference, your corrected profile, and reviewed_anatomy=true. This produces a separate, private draft containing:

• Skeleton
• Smoothed vertex weights
• Starting clips — grounded leg chains get walk/run with IK
• A GLB
• Editable Blender source

No step in this process auto-approves the result visually. Setting reviewed_anatomy=true records that you reviewed the anatomy — it does not mean the tool re-checked your correction for you.

Step 7 — Review deformation and starting clips

Open the Blender source or the GLB and check deformation on at least the starting clips: does a grounded leg's IK chain keep the foot planted without popping, does a spine segment twist instead of snapping, does a wing fold along its actual joint. Do not label this draft game-ready, and do not expect finger articulation, a facial rig, or finished combat animation — none of those are produced by this step.

Checklist

Troubleshooting

The uncertainty markers are everywhere, even on simple limbs. The source mesh likely has a pose, occlusion, or symmetry problem the four-view capture couldn't resolve. Check the source and annotated views, correct well-supported profile errors, and re-analyze when the body plan cannot be identified reliably.

Walk/run clips look wrong after rigging. Check whether the leg chain was actually classified as grounded in Step 3. Also check joint placement, floor contact and the job's report. Non-grounded leg chains may be omitted from locomotion generation; classification alone does not guarantee a correct walk cycle.

Corrected joints look fine in the profile but wrong in the rig draft. Confirm your edits were made in the coordinate frame returned with the profile, not a frame you reconstructed from the raw mesh transform.

FAQ

Does a successful analyze job mean the anatomy is correct?

No. It means a profile and four-view preview were generated; you still have to inspect it and, if needed, correct it before rigging.

Will the rig include finger or facial bones?

No. Facial and finger rigging are not produced by this workflow.

Can I get finished combat animation from the rig draft?

No. Grounded leg chains receive walk/run starting clips with IK; combat choreography is a separate, later effort.

Is the rig draft the same asset as my original creature mesh?

No. gripforge_creature_rig creates a separate private draft — skeleton, weights, starting clips, GLB, and editable Blender source — distinct from the analyzed source.

For the underlying tool contract, see the MCP documentation. For a broader look at what "game-ready" does and doesn't mean for a generated asset, see AI 3D model generator vs. game-ready asset. If you're setting up MCP tooling for the first time, MCP for game development with Claude Code and Cursor covers connection basics.