Games / AI games

AI games, without the slop

"AI game" has come to mean two opposite things: a game about AI-generated content, usually bad — or a game built with AI tools by people who care, which is just… game development now. Plops games are the second kind. This page is how that actually works, with receipts.

What "made with AI" means here

Every game in the Plops catalog is built in long working sessions between a very small human team and AI coding models (Anthropic's Claude family — we wrote up the specifics for Fable 5 and Opus). The model writes most of the code. The human decides what the game is, plays every build, argues about game feel, and reads what ships. It's the power-tool relationship: the saw cuts, the carpenter builds.

What it doesn't mean

• No AI-generated asset dumps. Plops ships zero image assets — every unit, glyph and effect is procedural canvas code. The art style is consistent because it's a codebase, not a folder of generated pictures that almost match.
• No prompt-to-game templates. These are server-authoritative real-time multiplayer games: 20Hz simulations, client prediction, reconnect handling, rating-based matchmaking, anti-cheat by architecture. That's engineering, and AI models are good at engineering when someone holds the design line.
• No unread code in production. The same rule as Plops Forge, our tip-based hub for AI tools and workflows, whose entire premise is "read the source before you run it." We apply our own rule to ourselves.

Why AI makes a two-person studio possible

A browser MOBA with ranked 2v2, guilds, payments, and cross-platform touch controls used to be a 30-person, two-year project. Plops is live, and its siblings (Toots, a 64-player wind-powered arena; Paintball; Power Kittens, built with actual small children as the QA team) are shipping from the same tiny room — because the model handles volume while the humans handle taste. The bottleneck stops being typing speed and becomes judgment, which is a much better bottleneck.

The honest costs

It's not magic. The model will happily build the wrong thing beautifully; someone has to notice. Game feel — the half-second of hitstop, the camera lag, whether a dash reads as juicy or floaty — still takes a human playing on a phone at midnight going "no, again." And multiplayer bugs remain multiplayer bugs: the model finds them faster, but the 2 a.m. deploy is still a 2 a.m. deploy. We publish the interesting parts on the devlog.

Go deeper

• Building a game with Claude Fable 5 — what changed with the newest model generation.
• Building a game with Claude Opus — the workhorse sessions that built most of Plops.
• Plops Forge — where we share the AI tools and workflows themselves, tip-based, source-first.