Lab
Experiments and side projects. Nothing here is a Semper Phoenix app.
Every time a new AI model comes out, people share videos of it building a whole game from a single prompt. I wanted to see that for myself, so most games here were made that way, either in one pass or by a team of AI agents working for hours. Others started from a prompt and then went back and forth with me until they felt right. Each card notes which AI model built it.
One prompt, a team of agents
One prompt each, then Claude's Ultracode mode had a team of AI agents design, build and test the game over many hours. I didn't change a line.
Bulwark
Turn-based tactics in a harbor city under siege. Sea monsters show every attack a turn ahead, so shove them into each other or the sea first. A solver proved every battle winnable.
One prompt, one pass
One prompt each. The AI wrote and tested the code on its own in a single pass, and I didn't change a line.
Moving Day
A truck-packing puzzle. Pick the smallest truck that fits, drag in the furniture, and load heavy things up front for three stars. A solver proved every move fits.
Delve
A turn-based dungeon crawler. Fight down ten generated floors and take the Heart of the Deep from its dragon. One life.
Tonegrid
A music toy. Paint notes and drums on a grid and they loop. Share a link so someone else hears your beat.
Gambit
Chess against an engine Claude wrote from scratch, with four strength levels, hints, and takebacks.
Nova Pinball
A neon pinball table with real ball physics, drop targets, multiball, and a dot-matrix score display.
Breach
A retro first-person shooter. Fight down through the generated sublevels of an overrun research facility to the exit.
Sandfall
A falling-sand playground. Paint with sand, water, oil, lava, acid, and gunpowder, then watch them react.
Blockworld
A block-building sandbox in an endless generated world with caves, forests, oceans, snowy peaks, and TNT. Your world saves in your browser.
Neon Swarm
A neon arena survival game. Your ship fires on its own while you dodge the swarm, collect energy, and pick upgrades.
Made together
These started from a prompt too, then I played them and asked for changes over several rounds. The AI still wrote all the code.
Fizzfin
An underwater bubble-pusher. Launch bubbles to squish piranhas, eels, and sharks, line up the gems, and spell PEARL for an extra life.
About the Lab
Every game started from a prompt to Claude, Anthropic's AI model, and Claude wrote and tested all of the code; I didn't edit any of it. For the one-prompt games, Claude chose the details on its own, and each card's build time runs from my prompt to a finished game that Claude had already tested itself, taken from the session log. When an automated code review later caught a real bug, Claude fixed that too, and nothing else changed.
The made-together games took more than one prompt. I played each version and asked for changes, so they have no single build time.
It was built in Claude's Ultracode mode. Instead of one pass, Claude ran 62 AI agents over about 19 hours: some designed the rules, some wrote the code, and others wrote 421 automated tests and tried to break it. A solver proved that every battle can be won without the city losing a single light, and the same solver powers the game's Show me button. It was still one prompt, and I didn't change a line.
No. Scores, saved games and settings stay in your own browser's storage. Nothing is sent to us or anyone else.
Yes. New experiments will show up here as I try new ideas and new models.