Visual Scene

Airplane Battle Challenge

A vertical shooter case for enemy waves, projectile patterns, hit detection, and readable survival feedback.

GameCanvasShooterProjectile collisionEnemy waves

Airplane Battle comparison episode poster

Watch episode

Published video

Airplane Battle comparison episode

This provisional comparison episode stitches together the current GLM, Doubao, Qwen, and DeepSeek runs recorded under the Claude Code framework.

2m 00s

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	GLM	Playable	A-83	Runs without fatal errors18/20 Core functionality25/30 Combat reliability17/20 Interaction quality13/20 Code sanity10/10	Review pending	unknown	Its combat presentation is a bit less dense and flashy than Doubao, so the lead comes more from stronger code quality and state coverage than from a clear gameplay blowout.
Claude Code	Doubao	Playable	A-82	Runs without fatal errors18/20 Core functionality26/30 Combat reliability18/20 Interaction quality16/20 Code sanity4/10	Review pending	unknown	The gameplay evidence is strong, but the reduced code-sanity score keeps it just behind GLM, and the current capture still leaves restart depth and broader state coverage less certain.
Claude Code	Qwen	Partial	B-60	Runs without fatal errors15/20 Core functionality19/30 Combat reliability10/20 Interaction quality10/20 Code sanity6/10	Review pending	unknown	It covers the basic loop, but enemy pressure, progression depth, and feedback richness are clearly lighter than the top two runs, so it does not yet read as equally complete.
Claude Code	DeepSeek	Partial	D45	Runs without fatal errors13/20 Core functionality14/30 Combat reliability7/20 Interaction quality7/20 Code sanity4/10	Review pending	unknown	It reaches a visible loop, but the menu, HUD density, progression, and overall feedback are the thinnest in the set, which keeps it well behind the stronger runs.

Claude Code

GLM

A-

Status: Playable
Score: 83
Breakdown: Runs without fatal errors18/20
Core functionality25/30
Combat reliability17/20
Interaction quality13/20
Code sanity10/10
Time: Review pending
Failure: Its combat presentation is a bit less dense and flashy than Doubao, so the lead comes more from stronger code quality and state coverage than from a clear gameplay blowout.

Claude Code

Doubao

A-

Status: Playable
Score: 82
Breakdown: Runs without fatal errors18/20
Core functionality26/30
Combat reliability18/20
Interaction quality16/20
Code sanity4/10
Time: Review pending
Failure: The gameplay evidence is strong, but the reduced code-sanity score keeps it just behind GLM, and the current capture still leaves restart depth and broader state coverage less certain.

Claude Code

Qwen

B-

Status: Partial
Score: 60
Breakdown: Runs without fatal errors15/20
Core functionality19/30
Combat reliability10/20
Interaction quality10/20
Code sanity6/10
Time: Review pending
Failure: It covers the basic loop, but enemy pressure, progression depth, and feedback richness are clearly lighter than the top two runs, so it does not yet read as equally complete.

Claude Code

DeepSeek

Status: Partial
Score: 45
Breakdown: Runs without fatal errors13/20
Core functionality14/30
Combat reliability7/20
Interaction quality7/20
Code sanity4/10
Time: Review pending
Failure: It reaches a visible loop, but the menu, HUD density, progression, and overall feedback are the thinnest in the set, which keeps it well behind the stronger runs.

What this tests

The hidden engineering inside a simple game

Movement and controls

Handles directional aircraft movement and firing across supported inputs.

Projectile systems

Maintains reliable player bullets, enemy bullets, and hit detection.

Enemy wave logic

Spawns and advances enemies in a way that creates a real combat rhythm.

Game-state management

Keeps score, damage, game over, and restart transitions coherent.

Visual readability

Makes aircraft position, incoming fire, hits, and danger states easy to see.

Debug iteration

Requires the agent to fix visible gameplay bugs through browser feedback.

Unified prompt

Build a playable browser-based airplane battle game. It should include a controllable player aircraft, enemy waves, projectile firing, projectile collision, score or survival feedback, player damage or lives, game over, and restart. Support keyboard controls on desktop and pointer or touch controls on mobile where appropriate. The final result must render visibly, avoid fatal runtime errors, and reset cleanly after restart.

Acceptance criteria

The app starts without fatal runtime errors.
The player aircraft, background playfield, and enemy units render visibly.
The player aircraft can move and fire with responsive controls.
Enemy units spawn and create a real playable combat loop.
Projectiles collide with enemies, the player, or required boundaries correctly.
Score, health or lives, game over, and restart logic work.
Restart clears stale state and starts a fresh playable round.
The final page does not produce console errors during normal play.

Scoring

Readable rubric for video and reproduction

Runs without fatal errors

The app starts, renders, and can be interacted with.

Core functionality

Movement, firing, enemy spawning, collision handling, and restart work.

Combat reliability

Bullets, hits, enemy waves, and player damage behave consistently.

Interaction quality

Controls feel usable and the battle remains readable under pressure.

Code sanity

The implementation is understandable and avoids obvious brittle shortcuts.

Run details

What each agent got right and wrong

GLM gameplay clipWatch clip

Screenshots

Menu Progress Game over

Claude Code

GLM

A-83

Runs without fatal errors18/20

Core functionality25/30

Combat reliability17/20

Interaction quality13/20

Code sanity10/10

Status: Playable
Strength: The run shows a full menu, boss-like encounter, readable HUD, explicit difficulty state, game-over flow, and strong audio presence.
Failure: Its combat presentation is a bit less dense and flashy than Doubao, so the lead comes more from stronger code quality and state coverage than from a clear gameplay blowout.

Doubao gameplay clipWatch clip

Screenshots

Menu Progress Game over

Claude Code

Doubao

A-82

Runs without fatal errors18/20

Core functionality26/30

Combat reliability18/20

Interaction quality16/20

Code sanity4/10

Status: Playable
Strength: This is the richest visible run so far: clear HUD, dense projectile play, pickups, combo feedback, and a complete game-over state.
Failure: The gameplay evidence is strong, but the reduced code-sanity score keeps it just behind GLM, and the current capture still leaves restart depth and broader state coverage less certain.

Qwen gameplay clipWatch clip

Screenshots

Menu Progress Game over

Claude Code

Qwen

B-60

Runs without fatal errors15/20

Core functionality19/30

Combat reliability10/20

Interaction quality10/20

Code sanity6/10

Status: Partial
Strength: The run includes a readable title screen, visible combat entities, scoring, weapon state, and a clean game-over transition.
Failure: It covers the basic loop, but enemy pressure, progression depth, and feedback richness are clearly lighter than the top two runs, so it does not yet read as equally complete.

DeepSeek gameplay clipWatch clip

Screenshots

Menu Progress Game over

Claude Code

DeepSeek

D45

Runs without fatal errors13/20

Core functionality14/30

Combat reliability7/20

Interaction quality7/20

Code sanity4/10

Status: Partial
Strength: The run reaches a visible combat state with working bullets, enemies, game over, and noticeable audio feedback.
Failure: It reaches a visible loop, but the menu, HUD density, progression, and overall feedback are the thinnest in the set, which keeps it well behind the stronger runs.

Postmortem

Why Airplane Battle Works As A Visual Arena Case

Projectile density stresses logic quickly

A shooter case forces repeated collision checks, wave timing, and damage handling under constant motion.

Readable failure is easy to capture

Broken enemy spawns, invisible bullets, unfair hitboxes, or bad restart handling are obvious in short clips and screenshots.

The draft is ready for real evidence

This page is a scaffold for run summaries, clips, screenshots, logs, and the final comparison episode.

Case package

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

Manifest Prompt Scoring Runs JSON

Commands

npm install
npm run dev
npm test

Use this scaffold to collect prompt, scoring, screenshots, clips, logs, and final run summaries before marking the case as published.

Airplane Battle Challenge

Airplane Battle comparison episode

One task, 4 recorded outcomes

Claude Code

Claude Code

Claude Code

Claude Code

The hidden engineering inside a simple game

Movement and controls

Projectile systems

Enemy wave logic

Game-state management

Visual readability

Debug iteration

Readable rubric for video and reproduction

Runs without fatal errors

Core functionality

Combat reliability

Interaction quality

Code sanity

What each agent got right and wrong

Claude Code

Claude Code

Claude Code

Claude Code

Why Airplane Battle Works As A Visual Arena Case

Projectile density stresses logic quickly

Readable failure is easy to capture

The draft is ready for real evidence

Ready for the next visual task?