I had GPT-6 Astra build Catan from a single prompt. One night later, the result was close enough to the real board game that I had to look twice. The prompt is below, word for word.
The prompt matters. The setup matters more. Read that part first.
The setup — read this first
- Open Codex. It lives inside ChatGPT. This is a build job, not a chat, so don’t paste this into a normal conversation.
- Set the model to GPT-6 Astra.
- Set the effort level to Max (see below).
- Paste the prompt exactly as written. Don’t shorten it, don’t “improve” it.
- Let it run. This is an overnight job — start it, walk away, check in the morning.
High vs Max. High will get you a playable build. Max is what built mine — it gives the model more room to plan, build and test before it hands anything back, and on a project like this the difference lands exactly where it matters: the rules edge cases, the opponents’ play, the polish that makes it feel like the real board instead of a demo. If you’re only going to run it once, run it on Max.
đź“‹ The prompt
Copy everything in the box and paste it as your first message.
Create a high-fidelity, fully playable single-player recreation of Catan that prioritizes authenticity to the original board game above all else.
Replicate the classic base game exactly: the 19-hex island with its standard terrain and number-token layout, the surrounding sea and harbors, the robber, and the full set of resource and development cards. Render it as a polished, illustrated top-down board with art direction, lighting and detail comparable to the real thing — hexes, roads, settlements, cities and cards should all read instantly as Catan.
Implement the complete official rules: snake-order initial placement, the distance rule, dice rolls and resource distribution, the robber and the discard rule on a 7, correct building costs, the development card deck with its play restrictions, Longest Road and Largest Army with correct transfer, harbor and maritime trading, player-to-player trading with counter-offers, and the first-to-10-victory-points win condition including hidden points.
Support a four-player game against three competent AI opponents with distinct strategies that place well, trade sensibly, target the robber deliberately, and race for Longest Road and Largest Army rather than playing randomly.
The interface must be a native, polished game UI: a clear turn structure, obvious feedback on valid build spots, an always-visible hand and resource panel, a build/trade/development action bar, an event log, and a scoreboard that tracks public and hidden points correctly. Prioritize fluid, self-explanatory interactions and an overall experience that feels like playing the real game rather than a simplified prototype.
Build it as a complete, immediately playable prototype that demonstrates deep rules fidelity and aesthetic fidelity.
Once it’s done
- Tell it to launch the game and play a full round. Judge it from the screen, not the code.
- Something off? Say it in one plain sentence — “the robber isn’t blocking resources”, “the bots never accept a trade” — and let it fix it. You don’t need to know why it broke.
- Want six players, a bigger island, house rules? Ask. You’re the game designer now.