Four-Player Battledrones: One Prompt, One New Multiplayer Map

Simon Willison released battledrones v14, which adds a map called Compass Rose and lifts the multiplayer limit from two players to four. The release itself is modest. The prompt behind it is worth reading, because it is a compact example of how to brief a coding agent on a feature that touches both content and networking.

What changed in v14

Per the release note, previous Battledrones maps were limited to two players. Compass Rose is a new map for four-player battles, and it works with the project's WebSocket-based, non-audio multiplayer mode. Up to four human players can join. According to Willison's note, Claude Opus 5.5 created the map and changed the multiplayer network code to support four players on it. Players join via share links or QR codes.

Anatomy of the prompt

Willison quotes the initial prompt he used. Read as a spec, it makes four separate requests:

• A design constraint: four-way rotational symmetry. This is a fairness property that can be checked, not a matter of taste.
• An AI requirement: up to four AI players.
• A transport requirement: it must work with the existing WebSocket, non-audio multiplayer mode, so the agent extends the current system instead of inventing a new one.
• A fallback rule: up to four humans can join, and any spare slots use AI.

The slot-filling rule is the one that matters most. It turns a hard-coded two-player assumption into a general case: a lobby of N humans fills the remaining 4 minus N seats with bots. Games with a fixed player count tend to bake that count into many small places, and asking the agent to generalise it is where a one-prompt change could go wrong.

Why this matters to builders

Three habits stand out, and they apply well beyond games.

1. Name the invariant. Rotational symmetry gives you something to verify, by eye or by test, after the agent finishes.
2. Anchor to existing architecture. Pointing the model at the current WebSocket mode keeps the diff inside code you already understand.
3. Specify the degenerate cases. Zero humans, one human and four humans all get covered when you say what happens to empty slots.

The same pattern applies to trading and fintech tooling, where multi-party state, such as several clients on one session or several agents on one book, is exactly the kind of thing that hides assumptions about N. A prompt that states the invariants and the fallback behaviour is easier to review than one that just says to add support for more players.

What to keep in mind

The release note describes what the agent produced, not how thoroughly it was tested. If you copy this workflow, review the network changes yourself. Check slot assignment, joining and leaving mid-game, and what the AI players do when a human disconnects. A symmetric map is only fair if the networking treats all four seats the same way.