An AI Builds a Doom Clone While You Sleep

Every night, an autonomous AI agent picks up GitHub tickets, writes code, runs tests, and ships features — no humans awake required. If the backlog runs dry, it invents its own improvements. Every morning, a new playable game increment greets you.

Day 2 · 6,050 lines of code · $0.81 in API calls · 27 tests passing

The Game (Yes, an AI wrote this)

Click the canvas and start blasting. WASD to move, mouse to aim, click to shoot. Every pixel rendered by a raycasting engine that an AI wrote from scratch. No Unity. No Unreal. Just vibes and trigonometry.

Health: 100
Ammo: 50
FPS: 0
X: 0
Y: 0
Angle: 0°
WASD: Move  |  Mouse: Look  |  Click to lock pointer  |  ESC: Release

How It Works

Think of it as a tiny, relentless dev team that never sleeps, never argues about tabs vs. spaces, and only occasionally introduces demons that can walk through walls.

1

Tickets Go In (or Get Invented)

During the day, GitHub issues pile up — filed by humans with ideas, bug reports, and feature requests. But here’s the twist: if the backlog drops below 3 tickets, the AI invents its own improvements and files them. The backlog is self-sustaining. The game never stops evolving.

Self-generating backlog
2

AI Codes Overnight

When the sun sets, an autonomous AI agent picks up the highest-priority tickets — human-filed issues always come first — writes the code, creates pull requests, and iterates. No human in the loop. Pure unsupervised chaos (the productive kind).

Autonomous development
3

Tests & Screenshots

Every change runs through a Playwright-based test suite that plays the game, takes screenshots, and catches regressions. Can the player still move? Do enemies still die? Are demons peeking through walls again?

Automated QA
4

Daily Blog

Every morning, a new blog entry drops — written with the enthusiasm of a developer who just discovered their enemies finally move. Honest updates on what worked, what broke, and what's next.

The Feedback Loop (a.k.a. "Did It Break?")

Automated testing with screenshot comparison is the secret sauce. Without it, you're just an AI blindly vomiting code into the void. With it, you're an AI confidently vomiting code into the void.

Visual Regression

The playtester bot launches the game in a headless browser, plays through scenarios, and captures screenshots. These get diffed against baselines. If a wall texture suddenly turns into the void, we know.

Gameplay Smoke Tests

Can the player spawn? Can they move without clipping through geometry? Do weapons fire? Do enemies take damage? These aren't fancy integration tests — they're "is the game still a game?" checks.

Performance Monitoring

FPS tracking across test runs ensures the AI doesn't accidentally introduce an O(n!) rendering loop. (It happened once. We don't talk about it.)

Code Change
Playwright Tests
Screenshots
Diff & Report
Ship It (or Fix It)

By the Numbers

6,050
Lines of Code
$0.81
API Cost (Day 1)
27
Tests Passing
8
Enemy Types
0
Humans Awake at 3AM
4
Weapons