Visual Scene

Magician Card Toss Scene

Four AI models animated the same stage magician prompt. The strongest run was not the busiest one, but the one with the clearest hand-to-card causality.

Visual sceneAnimationStagePlaying cardsMotion logic

Final four-model magician comparison poster

Watch episode

Published video

Final four-model magician comparison

Full landscape comparison of the four magician card-toss runs before short-form cuts.

0m 22s

Watch full episode Watch comparison short Watch best-run short Watch worst-run short

Result matrix

One task, 4 recorded outcomes

The table shows visible results first. Detailed logs and clips can be attached once real runs are recorded.

Agent	Model	Status	Score	Breakdown	Time	Cost	Failure
Claude Code	Opus	Strong	A-85	Visual authenticity (VQ)34/40 Physics and consistency (PC)20/25 Semantic restoration (SR)17/20 Composition and readability (CR)14/15	22s clip reviewed	unknown	The hand-to-hand card path is still somewhat conservative, and the scatter effect appears late rather than dominating the loop.
Claude Code	GLM	Solid	B+77	Visual authenticity (VQ)32/40 Physics and consistency (PC)16/25 Semantic restoration (SR)16/20 Composition and readability (CR)13/15	22s clip reviewed	unknown	The card motion is often busy rather than physically grounded: several cards appear near the head or above the body without clearly originating from the throwing hand.
Claude Code	Qwen	Partial	B72	Visual authenticity (VQ)29/40 Physics and consistency (PC)16/25 Semantic restoration (SR)15/20 Composition and readability (CR)12/15	22s clip reviewed	unknown	The card paths read more like vertical drift or isolated tosses than a continuous left-hand-to-right-hand exchange, and the fan scatter is less explicit.
Claude Code	Doubao	Weak	C-46	Visual authenticity (VQ)18/40 Physics and consistency (PC)10/25 Semantic restoration (SR)9/20 Composition and readability (CR)9/15	22s clip reviewed	unknown	The scene is too sparse for the prompt: card motion mostly stays near the body, the hand-to-hand throw is unclear, and the upward fan scatter is weak.

Claude Code

Opus

A-

Status: Strong
Score: 85
Breakdown: Visual authenticity (VQ)34/40
Physics and consistency (PC)20/25
Semantic restoration (SR)17/20
Composition and readability (CR)14/15
Time: 22s clip reviewed
Failure: The hand-to-hand card path is still somewhat conservative, and the scatter effect appears late rather than dominating the loop.

Claude Code

GLM

B+

Status: Solid
Score: 77
Breakdown: Visual authenticity (VQ)32/40
Physics and consistency (PC)16/25
Semantic restoration (SR)16/20
Composition and readability (CR)13/15
Time: 22s clip reviewed
Failure: The card motion is often busy rather than physically grounded: several cards appear near the head or above the body without clearly originating from the throwing hand.

Claude Code

Qwen

Status: Partial
Score: 72
Breakdown: Visual authenticity (VQ)29/40
Physics and consistency (PC)16/25
Semantic restoration (SR)15/20
Composition and readability (CR)12/15
Time: 22s clip reviewed
Failure: The card paths read more like vertical drift or isolated tosses than a continuous left-hand-to-right-hand exchange, and the fan scatter is less explicit.

Claude Code

Doubao

C-

Status: Weak
Score: 46
Breakdown: Visual authenticity (VQ)18/40
Physics and consistency (PC)10/25
Semantic restoration (SR)9/20
Composition and readability (CR)9/15
Time: 22s clip reviewed
Failure: The scene is too sparse for the prompt: card motion mostly stays near the body, the hand-to-hand throw is unclear, and the upward fan scatter is weak.

What this tests

The hidden engineering inside a simple game

Character staging

Keeps the magician centered, readable, and dressed according to the prompt.

Card motion logic

Shows whether card paths originate from the hands and preserve believable hand-to-card causality.

Secondary action

Adds a periodic upward toss and fan-shaped scatter without breaking the main hand-to-hand card loop.

Stage composition

Uses theatrical curtains, floor, spotlight, and hierarchy to make the scene instantly readable.

Unified prompt

Create a single-file HTML page (including HTML, CSS, and JavaScript). In the center of the page, place a stylized magician wearing a black tailcoat and a top hat. The magician continuously throws playing cards from his left hand to his right hand, with the cards spinning and flying through the air. At the same time, every few seconds, he randomly tosses a card upward, creating a fan-shaped scattering effect. The background should be a theatrical stage with a spotlight. All animations should run smoothly in a continuous loop.

Acceptance criteria

The result is a single-file HTML scene.
A stylized magician is centered in frame.
The magician wears a black tailcoat and a top hat.
Cards visibly travel from the left hand to the right hand.
Cards spin or rotate while flying through the air.
A periodic upward toss creates a fan-shaped card scatter.
The background reads as a theatrical stage with a spotlight.
The animation loops smoothly without obvious jumps or broken state.

Scoring

Readable rubric for video and reproduction

Visual authenticity (VQ)

Judges magician silhouette quality, costume finish, card rendering, lighting, stage design, and overall visual polish.

Physics and consistency (PC)

Judges whether card paths plausibly originate from the hands, whether left-to-right throws read as actual hand-card interaction, and whether the upward fan scatter has a believable cause.

Semantic restoration (SR)

Judges whether the result covers the full prompt: magician, black tailcoat, top hat, spinning cards, hand-to-hand throwing, upward fan scatter, stage, spotlight, and smooth loop.

Composition and readability (CR)

Judges subject clarity, hierarchy, card visibility, and whether the viewer can understand the magic-card action immediately.

Run details

What each agent got right and wrong

Opus landscape clipWatch clip

Screenshots

Captured run

Claude Code

Opus

A-85

Visual authenticity (VQ)34/40

Physics and consistency (PC)20/25

Semantic restoration (SR)17/20

Composition and readability (CR)14/15

Status: Strong
Strength: Best overall run: the stage, spotlight, magician silhouette, and card scatter are clean, and the card motion reads more plausibly connected to the hands than the busier alternatives.
Failure: The hand-to-hand card path is still somewhat conservative, and the scatter effect appears late rather than dominating the loop.

GLM landscape clipWatch clip

Screenshots

Captured run

Claude Code

GLM

B+77

Visual authenticity (VQ)32/40

Physics and consistency (PC)16/25

Semantic restoration (SR)16/20

Composition and readability (CR)13/15

Status: Solid
Strength: Most visually active card field in the set, with a complete stage, centered magician, and many readable cards in flight.
Failure: The card motion is often busy rather than physically grounded: several cards appear near the head or above the body without clearly originating from the throwing hand.

Qwen landscape clipWatch clip

Screenshots

Captured run

Claude Code

Qwen

B72

Visual authenticity (VQ)29/40

Physics and consistency (PC)16/25

Semantic restoration (SR)15/20

Composition and readability (CR)12/15

Status: Partial
Strength: Readable centered magician, strong spotlight framing, and visible card tosses that make the stage-magic intent clear.
Failure: The card paths read more like vertical drift or isolated tosses than a continuous left-hand-to-right-hand exchange, and the fan scatter is less explicit.

Doubao landscape clipWatch clip

Screenshots

Captured run

Claude Code

Doubao

C-46

Visual authenticity (VQ)18/40

Physics and consistency (PC)10/25

Semantic restoration (SR)9/20

Composition and readability (CR)9/15

Status: Weak
Strength: The central magician and spotlight are immediately readable, and the simplified silhouette avoids visual clutter.
Failure: The scene is too sparse for the prompt: card motion mostly stays near the body, the hand-to-hand throw is unclear, and the upward fan scatter is weak.

Postmortem

Why The Busiest Magician Was Not The Best Run

Card quantity is not card logic

GLM puts many cards on screen, but the strongest interpretation is the one where cards appear to come from the magician's hands and preserve a believable cause-and-effect relation.

The second action matters

The prompt asks for both a continuous hand-to-hand throw and a periodic upward fan scatter. Weak runs either underplay the scatter or make it appear disconnected from the character.

Stage readability helps, but motion decides the ranking

All four runs establish some version of a magician under a spotlight. The ranking spread comes from how clearly each one animates the cards as an intentional magic trick rather than decorative particles.

Case package

starter/
tests/
prompt.md
task.md
scoring.md
runs.json
screenshots/
clips/
logs/

Manifest Prompt Scoring Runs JSON

Commands

npm run capture:screenshots -- <html-files...> --case magician --output-dir images
npm run capture:videos -- <html-files...> --case magician --output-dir videos
npm run validate

This package stores the prompt, scored runs, screenshots, and local clips used to generate the full comparison and short-form cuts.

Magician Card Toss Scene

Final four-model magician comparison

One task, 4 recorded outcomes

Claude Code

Claude Code

Claude Code

Claude Code

The hidden engineering inside a simple game

Character staging

Card motion logic

Secondary action

Stage composition

Readable rubric for video and reproduction

Visual authenticity (VQ)

Physics and consistency (PC)

Semantic restoration (SR)

Composition and readability (CR)

What each agent got right and wrong

Claude Code

Claude Code

Claude Code

Claude Code

Why The Busiest Magician Was Not The Best Run

Card quantity is not card logic

The second action matters

Stage readability helps, but motion decides the ranking

Ready for the next visual task?