← All posts

Battle Drones: Simon Willison's Claude Opus 5.5 RTS game

2026-10-20 · 4 min read

Battle Drones is a free real-time strategy game that runs in your browser, and Simon Willison built it almost entirely with Claude Opus 5.5 and Claude Code. You can play it right now at simonw.github.io/battledrones on a phone or a laptop without installing anything.

Willison announced it on X today as a "Command & Conquer style RTS called Battle Drones, mainly using Claude Opus 5.5". In a follow-up post he added that "Claude generated all of the 3D models (in Three.js)", and he linked a model viewer where you can look at every unit and its animations. His favorite is a giant stomping mech called the Goliath.

The game is fun on its own, but I found the build story more interesting, mostly because of how little of the code he read.

What Battle Drones is

According to Willison's write-up, the game has a skirmish mode, a Last Stand survival mode, a Daily Op challenge, an eight-mission campaign, and an AI vs. AI mode. It also has fog of war, six maps across five biomes, a random map generator, and "agent battles" where language models play each other. His example match pits Claude Haiku 5.5 against GPT-6 Luna, with a replay you can step through.

The whole thing is static HTML and JavaScript. There's no server behind it, and the game works offline as a web app.

How he built it

Willison says he has "reviewed almost none of the code." The project is written in TypeScript, a language he has barely used. Coding agents, mostly Claude Code, made 321 commits across 12 major versions. The first version, from July, was built with Claude Fable 5, and Opus 5.5 took over from version 3.

The early versions worked, but they weren't much fun. In his words, v1 "wasn't any fun to play." Version 4 started from a short prompt:

This game isn't fun enough yet. Make it a LOT more fun.

Claude then ran more than a thousand simulated AI vs. AI matches and used the results to rebalance units and add features such as opening choices and call-down powers. His release notes give numbers: games still undecided after 25 minutes fell from 14% to 0.9%.

Later versions followed the same pattern. A single prompt ported the whole game to 3D in Three.js. For the bigger features, he wrote a design prompt and asked Claude to research and critique the idea before it built anything. The long runs used dynamic workflows in Claude Code, where one script coordinates many subagents. Anthropic has since brought the same feature to its hosted agents, which we covered in Claude Managed Agents dynamic workflows. He also admits this mode "burns tokens like there's no tomorrow," and the runs often went all night.

How to try it

Open the game in any modern browser. Touch controls work on a phone. A few things worth trying:

  • Play a skirmish against the computer to learn the units.
  • Open the model explorer to see each 3D model Claude wrote in code, with its animation states.
  • Watch an AI vs. AI match if you'd rather see the balance work Claude did.
  • Try audio multiplayer, which links two devices within microphone range by sending the game state as sound. Willison warns on X that it plays a high-pitched noise the whole time and "might give you an actual headache."

If you liked the Opus 5.5 pixel art and animation posts from September, it's the same idea at a much bigger scale. The 3D models and the soundtrack in Battle Drones are code the model wrote, though the early 2D sprites came from OpenAI's image API.

My read

The prompt is the part I keep coming back to. "Make it a LOT more fun" is vague, and it worked because Claude could test its own changes against a measurable goal by playing a thousand games and checking the results. Willison didn't specify what fun meant. He gave the agent a loop it could run by itself.

There's a catch, though. He shipped code he hasn't read, and he says himself that games are "super low stakes." If a unit is overpowered, nobody gets hurt. Handing an agent the code behind your invoicing or your customer records without reviewing it would be a very different decision.

What this means for a business

Agents can now build something polished when they have a clear way to check their own work. In a business, that looks like reports reconciled against the books, or a lead form tested against real submissions. When nothing checks the work, you still need a person reading it. If you'd like help finding which of your processes have that kind of built-in check, New Face Design offers a free process audit.

08 / Start here

Which of this can you use today?

Tell us what you use today and we'll reply within one business day with what it would take. Free, 20 minutes, no pitch deck.

Email

pgorski@newfacedesign.com

Phone

+1 (773) 627-2176

Based in

Chicago area

Working with clients everywhere

Skip the form: grab a 20 minute slot.

Open the calendar →