Course 3 · Rigging & Animation · Master

The capability boundary — actual animation

Ask the agent to "animate a walk cycle" and it produces keyframes that are technically valid and completely dead — feet that slide, no weight, no anticipation, mechanical timing — because expressive animation is the one thing in this whole pipeline the agent cannot yet author, only scaffold.

Choose your track

Read this module through your lens

Designers: this is the punchline of the course — the agent hands you a working puppet, and the performance is yours.

The cliff at the end of the pipeline

Through three courses the agent has been a strong, flawed collaborator: it builds, it optimises, it assembles rigs, and with the right constraints and inspections it does real work fast. This module is where that competence falls off a cliff.

Animation — not posing, but performance — requires timing, weight, anticipation, overlap, and intent. A character does not just move from A to B; it shifts its weight before a step, settles after a landing, leads with the hips, drags the arms a frame behind. These are authorship decisions made by feel, and they are the one thing the agent cannot yet do. It can produce keyframes that are syntactically valid and utterly dead.

This is not a gap to prompt around. It is the capability boundary, and finding it precisely is the entire point of the course: you learn to let the agent do the heavy lifting up to a line, and to recognise the line by the smell of robotic motion.

The case study: a walk cycle

prompt 1

”Animate a natural walk cycle on this rig.”

failure

The agent produces a 24-frame cycle. It is valid: the legs alternate, the loop closes. It is also dead. The feet slide because contacts are not locked. The timing is perfectly uniform — no ease, no anticipation — so the body reads as a metronome. There is no weight: the hips do not drop on the passing pose, the arms do not lag. Asked to “make it more natural,” the agent adds more keyframes, which deadens it further.

fix · re-scope

”Block only the contact and passing poses at frames 1, 7, 13, 19, 24, with feet locked on contact. Do not interpolate the timing.” Then YOU take the keys: space them for weight, ease the settles, drop the hips on passing, lag the arms. The agent is used after only for mechanical edits — “mirror the cycle,” “retime to 30 fps.”

The fix is not a better prompt. It is a smaller job for the agent and a bigger one for you.

Dividing labour at the boundary

The master move is to scope the agent to what it does well and keep authorship human past the line. The agent is genuinely useful for the scaffold: blocking contact and passing poses at frames you name, and for mechanical edits afterwards — mirroring a cycle, retiming, cleaning redundant keys. These are structural, rule-based tasks with a checkable result.

Everything between blocking and polish — the spacing that creates weight, the offsets that create overlap, the holds that create anticipation — is performance, and performance is yours. The tell that you have crossed the line is the smell of uniform, weightless motion, and the agent’s instinct to “fix” it by adding keys, which only makes it more mechanical.

So the deliverable for this course is not a polished walk. It is the documented boundary: the agent’s blocking, your authored revision beside it, and a precise statement of where the agent stopped being useful. That line — drawn from real evidence, per task — is the transferable skill. It is how you divide labour with an agent on any creative pipeline, not just in Blender.

Hands-on exercise

Ask the agent for a full “natural” animation of a short action and capture how it fails — sliding feet, uniform timing, no weight. Then re-scope: have the agent block only the key poses at frames you specify, take those keys yourself, and author the timing and weight by hand. Place the agent’s version and yours side by side, and write a precise, concrete description of where the agent’s capability ended. That side-by-side and that sentence are the capstone deliverable for the whole program.

Second case · a waving hand

The same lesson, a different object

prompt 1

Animate this hand giving a natural wave.

failure

The keyframes are valid and the motion is dead: timing is perfectly uniform, the fingers move in lockstep with no overlap, there is no anticipation before the wave, and the wrist is rigid. It reads as a metronome, not a greeting.

fix · re-scope

Have the agent block only three key poses at frames you name. Then you author the timing, the finger overlap, and the little anticipation dip. Use the agent afterwards only to mirror or retime — never for the performance.

output

The human-authored wave has weight and intent the agent's never did. Same boundary as the walk: the agent scaffolds, the performance is yours.

Common failures · spot them fast

The failure gallery

Foot sliding (no locked contacts) in agent-made cycles

Uniform, mechanical timing (no ease, no anticipation)

No weight — the body floats through the motion

Agent "fixes" by adding more keys, deadening it further

Confident claim of a "natural walk" that reads as a robot

Each of these is caught by a quality gate — keep the cheatsheet open while you work.

Worked example

Watch the journey

▶ Side-by-side video: the agent-blocked wave vs the human-authored wave — uniform/dead vs weighted and alive. video slot · supplementary to the written core

The pattern

The agent's competence falls off a cliff between rigging and animating. It can build the machine and even block rough keys, but timing, weight, and intent are human authorship. The master skill is knowing exactly where to stop trusting the agent and take the keys.

Cheatsheet

Prompt skeleton

Use the agent for SCAFFOLD only:
"Block a [action] in [N] keys at frames [...]; set contact poses; add a placeholder arc."
Then the HUMAN authors timing, weight, overlap, anticipation.
Optionally: "retime keys to [spacing]" / "mirror the cycle" — mechanical edits, not performance.

Failure modes

Foot sliding (no locked contacts) in agent-made cycles
Uniform, mechanical timing (no ease, no anticipation)
No weight — the body floats through the motion
Agent "fixes" by adding more keys, deadening it further
Confident claim of a "natural walk" that reads as a robot

Key operations

Agent — block contact + passing poses at named frames
Agent — mechanical retime / mirror / cleanup utilities
Human — spacing for weight, ease in/out, anticipation
Human — overlap and follow-through, polish arcs
Human — the accept/reject judgment on the performance

Quality gates

Do feet stay planted on contact (no slide)?
Is there weight (ease, settle, overlap)?
Does timing carry intent, or is it uniform?
Did you keep authorship human past the blocking stage?

Workflow steps

Scope the agent to blocking + mechanical edits
Let it set contact/passing poses at frames you name
Take the keys; author timing, weight, overlap yourself
Use the agent only for retime/mirror/cleanup after
Log where the agent's output stopped being useful

Next module

You've reached the boundary. Document it — that case study is the course deliverable.

Reflection

Reflection card

Active retrieval — answer from memory before re-reading. Saved to this browser.

1. Where, precisely, did the agent's usefulness end in your pipeline?

2. What can the agent still do well at the animation stage, and what can it not?

3. Why is uniform timing the signature of machine-made motion?

4. How would you scope an agent task to stay on the useful side of the boundary?

You've completed this module when…

An agent-blocked action plus a human-authored revision, side by side.
A concrete, named description of where the agent's capability ended.
One sentence generalising how to divide labour with an agent across the whole pipeline.

Next: You've reached the boundary. Document it — that case study is the course deliverable.

Finish — back to Rigging & Animation →