GPT images 2.0 is a huge unlock for game developers
Visualize what the ideal game state is before even spending time on animation looking weird.
Storyboards this character scene, and mockup with GPT and animated it with grok.
Got Wesnoth (20 yrs old, 1M lines of C++) running natively in browser. Campaigns and multiplayer both work, including full wire protocol, saves persist.
Four source files changed, the other 999,996 lines are completely untouched.
~18s load time but worth the wait.
I needed 400 AI-generated game sprites for this game. All consistent and production-ready, generated overnight.
One right image takes an hour. A pipeline that does 400 correctly is something else entirely.
Here's every specific thing that broke, and exactly how I fixed it
#6: Six parallel agents, all writing to /tmp/nobg.png
One agent's output overwrote another's mid-process. Race condition and silent data corruption
Always use mktemp.
Bonus: checking for valid PNGs with grep "89 50 4e 47" against xxd output fails because xxd renders w/o spaces
Continuing work on this cinematic platformer.
Layered in multiplayer dynamics alongside @theworldlabs splats and insane challenges. Playable in browser and entirely vibe coded.
Core game loop + taste = great clicker games
Without taste no one is getting addicted to your game. With too much stylistic complexity no one will even try play it.
This hacking simulation clicker is a wonderful example.
Another stunning indie game thatโs entirely vibe coded and built with @threejs
Full of RPG-like quests, customization, and amazing in-game physics.
Discovered this one while diving deep into award-winning AI games in the @chromaawards archives.
This seriously deserves more appreciation: rosebud.ai/p/traveller-1
This indie racing game built with three.js is mindblowing. The audio effects make it peak gameplay even for a simple idea & unbelievable that itโs entirely vibe coded
I ported a 20-year old 1M-line C++ strategy game to run in your browser & it cost me $300. The second one was only $40.
The first port is expensive because you're building the workflow at the same time you're doing the work. Each one after that is quicker.
What I learned:
TLDR: encoding context and feedback into persistent artifacts that agents use across runs. The core idea is that the harness is doing memory work the context window can't.
@OpenAI calls this harness engineering
And now you can play on @Rosebud_AI
