I'm currently seeking feedback on the first iteration of my newest game, tentatively titled, "Yardwork." It's being built using Github Copilot and Phaser 3. So far, development has gone smoothly, but I know I still have more to learn about using Tiled and creating tiled sprites. Any feedback will be welcomed and appreciated!
Just wanted to let everyone know that my latest NES-inspired game, "Birdie Blues & the Jazz Cats," is now available on itch.io! Play for free today!
Every time someone posts an AI-assisted game, the comment section splits: "the AI made this, not you" vs "AI is just a tool, like a compiler." I've been on both sides of that argument. Now I've shipped something with Copilot as a real collaborator, and I want to give you something more useful than a take — specifics.
The game: Birdie Blues and the Jazz Cats — a free, browser-playable NES-style single-screen arcade platformer. The concept is a mashup of three classics: Donkey Kong's vertical maze structure, Bubble Bobble's "the threats are also your ammo" economy, and Popeye's two persistent, unkillable pursuers with distinct AI personalities. You're a bird. A manic Pianist at the top of the stage hurls music notes down at you. Walk into a note to catch it — that's your ammo. Press X to throw it at a Jazz Cat and knock it into a 5-second sleep. There's a Tempo Timer counting down. Six stages. They loop forever and get faster every lap.
Took about two weeks of actual development plus a few days of polish. Solo dev. My eighth game.
What Copilot actually did:
Architecture scaffolding — and it paid dividends. Early in the project I needed all cross-system communication (score, audio cues, life loss, achievement triggers) to route through named event constants rather than direct calls between entities. Copilot generated the structure for this — events.js as a single source of truth for event names, an event bus that everything subscribes to — cleanly and fast. The payoff: when I added achievements weeks later, they just listened for events that already existed. When audio hooks needed wiring, the connection points were already in place. This is the kind of architectural pattern that's easy to skip when you're moving fast. Copilot made it the default.
BFS pathfinding — the scaffold came fast, the feel took weeks. The Jazz Cats need to navigate multi-tiered platforms and ladders. Copilot helped me build the BFS pathfinding graph over each stage's platform and ladder layout quickly. What Copilot could not do: figure out the six tuning parameters per cat — speeds, commit windows, jitter, targeting offsets, lead distance — that make the cats feel threatening instead of annoying. That was entirely manual iteration. There was a point where I genuinely wasn't sure the game was going to work. No combination of parameters was landing. The AI had built me a functional pathfinder. It had no idea what "feels like it's hunting you" means.
Physics boilerplate — fast and clean. Coyote time, jump buffering, spawn invulnerability timing — Copilot produced tight first drafts for all of it. Minimal editing required. This is where AI assistance is clearly, unambiguously good: patterns that are well-understood in the abstract, annoying to re-type from scratch, and not the interesting part of the game.
Tests and config scaffolding. Vitest unit tests, tunables centralised in config.js, naming conventions for assets — Copilot generated these quickly and consistently. Less glamorous than the AI discourse suggests. More useful than the AI discourse acknowledges.
What Copilot couldn't do:
It couldn't tell me the cats needed a flanking personality in addition to a direct-chase personality. That insight came from replaying Popeye on an emulator until I understood why two pursuers feel different from one.
It couldn't tell me when the physics feel was right. Six weeks of NES muscle memory lives in my hands. A language model doesn't have a sense of when a jump arc is off.
It didn't produce the emergent flanking behavior that makes the two cats feel like they're cooperating without any cooperation code. Cat A (The Stalker) charges directly across every platform. Cat B (The Flanker) climbs ladders to cut off your exit. Neither has any awareness of the other. Together they create real pressure. That required understanding the design deeply enough to know what to tune, in what order, and why.
The honest version of the AI debate:
Copilot made me faster. It made the boring parts shorter. It never made a design decision. It never told me something was wrong. It generated a lot of code I deleted. The game is mine in every way that matters — every stage layout, every AI personality, every tuning call, every "this doesn't feel right, start over" moment.
It's also genuinely useful in ways I didn't fully appreciate until I shipped with it. The event bus would have been tempting to skip if I'd had to write it entirely from scratch. I didn't skip it. That made the last third of development significantly cleaner.
Play it free in your browser — no install, no account: https://orangeheartcreative.itch.io/birdie-blues
Drop your Stage and score in the comments. Can you clear Stage 3 on your first run?