Where auto-weights betray you
Auto-weights give the agent a green light it can't read — the model binds, the agent reports success, and then an arm bends and the shoulder collapses like a crushed can, because weight quality is only visible in a pose the agent never struck.
Choose your track
Read this module through your lens
Skinning is judged in motion, and the agent can’t move it well
Once bones sit in the joints, the mesh has to follow them — that is skinning, or “weights”: every vertex is assigned how strongly each bone pulls it. The agent can bind a mesh with automatic weights in one call and will report success. But automatic weights are only a starting guess, and their quality is invisible in the rest pose where the agent leaves the model.
The truth shows only under deformation. Bend an arm and a badly weighted shoulder pinches into a “candy wrapper”; a stray long-range weight drags distant vertices along; an unweighted vertex stays frozen in rest while its neighbours move, tearing the surface. All of this is fine until something rotates.
Here the course’s pattern sharpens into something stronger: the agent is not just under-instructed, it is structurally blind to the property being judged. Deformation quality requires posing and looking — a perceptual act the agent does not perform.
The case study: a character’s shoulder
”Skin this character mesh to the armature with automatic weights.”
”Mesh bound with automatic weights. Rig is ready to animate.” In rest pose, flawless.
You raise the arm 90 degrees. The shoulder collapses into a pinched crease, the chest near the armpit caves, and a few vertices on the side stay behind — they were left unweighted. The agent had no way to know: it never raised the arm, and even if it rendered one, it cannot perceive “this deformation is wrong.”
”Limit to 4 influences per vertex, normalise, mirror weights across the X axis, and smooth the weights around the shoulder. Report any unweighted vertices.” Then you weight- paint the armpit by hand and re-pose.
The human is the test harness
The agent can do real work here — bind, limit influences, normalise, mirror, smooth — and you should ask for all of it, because each removes a class of defect. But none of it substitutes for the test, and the test is you posing the rig.
So the workflow inverts the usual order: the agent acts, then you become the evaluation function. Take each joint through its full range — arm up, elbow bent, knee folded — and watch for pinches, bleed, and frozen vertices. When you find one, that specific area is where you (or a targeted weight-paint prompt) intervene. Then re-pose to confirm.
This is the clearest case so far of human-in-the-loop by necessity, not preference. Earlier modules asked the agent to report numbers you could check. Skinning asks you to supply a perception the agent does not have. Notice that shift — it is the whole theme of Course 3, and it intensifies in the final module.
Hands-on exercise
Bind a jointed mesh with automatic weights and confirm it looks perfect in rest pose. Then pose every joint through its range and hunt for the first deformation defect — pinch, bleed, or an unweighted vertex left behind. Screenshot the worst pose. Apply the weight fixes (limit influences, smooth, mirror, or hand-paint), re-pose, and screenshot the improvement. Write one sentence on why this step could not be left to the agent alone.
Second case · a single finger
The same lesson, a different object
Skin this finger mesh to its bones with automatic weights.
Bound with automatic weights. Rig is ready to animate. In rest pose, flawless.
Bend the knuckle and the joint pinches to a sharp crease, while a thin strip of vertices on the side stays frozen in rest — they were left unweighted. The agent never bent it, and could not have seen the pinch if it had.
Limit to 3 influences, normalise, smooth weights around the knuckle, and report unweighted vertices. Then hand-paint the frozen strip and re-pose to confirm.
Common failures · spot them fast
The failure gallery
Candy-wrapper pinch at joints when posed
A bone influencing far-away vertices (long-range bleed)
Unweighted vertices left behind (faces stay in rest pose)
Too many influences per vertex (mushy deformation)
Weights judged in rest pose, where everything looks fine
Each of these is caught by a quality gate — keep the cheatsheet open while you work.
Worked example
Watch the journey
The pattern
Cheatsheet
Cheatsheet
Prompt skeleton
Skin [mesh] to [armature] with automatic weights, then:
Limit to <= [4] influences per vertex; normalise weights.
Mirror weights across symmetry. Smooth weights at [joints].
I will pose and judge deformation; report influence counts + any unweighted vertices.
Failure modes
- Candy-wrapper pinch at joints when posed
- A bone influencing far-away vertices (long-range bleed)
- Unweighted vertices left behind (faces stay in rest pose)
- Too many influences per vertex (mushy deformation)
- Weights judged in rest pose, where everything looks fine
Key operations
- Bind with automatic/heat-map weights as a starting point
- Limit total influences per vertex
- Normalise + mirror weights across symmetry
- Smooth weights around problem joints
- Pose the rig to evaluate (the only real test)
Quality gates
- Does each joint deform smoothly when posed (no pinch)?
- Any vertices left unweighted (stuck in rest)?
- Influences per vertex within the limit?
- Symmetric weights on symmetric geometry?
Workflow steps
- Auto-bind, limit + normalise influences
- Pose each joint through its range
- Find the pinch/bleed; weight-paint or smooth that area
- Re-pose to confirm
- Log the worst pose before/after
Next module
- rig_ik_and_constraints — controllers, IK chains, and pole targets.
Reflection
Reflection card
Active retrieval — answer from memory before re-reading. Saved to this browser.
You've completed this module when…
- A bound mesh posed through at least one full joint range, screenshotted.
- One deformation defect (pinch, bleed, or unweighted vertex) identified in a pose.
- The weight fix and a re-posed screenshot showing improvement.
Next: rig_ik_and_constraints — controllers, IK chains, and pole targets.
Finish — back to Rigging & Animation →