Build a game with a coding agent: the stack, time and cost behind eight October 2026 games

The run: Battle Drones from Phaser sprites to a three.js RTS

Battle Drones is the only one of the eight games with its code, commits and most of its prompts public, so it is the run to copy. Fourteen versions went up between 17 July and 24 October 2026, most after 6 October, in 346 commits written mainly by Claude Code.

The Compass Rose map in Battle Drones v14, captured by the agent. Image: simonw/battledrones by Simon Willison, Apache-2.0, resized.

Fable 5 builds a game nobody wants to play

Version 1, in July, came from a design Q&A with Claude Fable 5. Willison's verdict: it "wasn't any fun to play". Version 3, now on Opus 5.5 in October, added maps and regenerated art and still "wasn't very fun".

One prompt and an overnight workflow add the fun

Version 4 came from "This game isn't fun enough yet. Make it a LOT more fun.", run as a Claude Code dynamic workflow from Willison's phone, from evening until "late into the next morning". Willison describes over a thousand simulated playtests; Claude's own release notes count 1,200 AI-against-AI test matches and say games still undecided after 25 minutes fell from 14% to 0.9%. Those are the agent's numbers, relayed by Willison and not checked by anyone else.

A port to three.js and a map per prompt

Version 5 was another overnight run: "Port this game to three.js with 3D models for all of the buildings and vehicles and landscapes". Later versions added units, an archipelago map from a prompt that included "First research and then critique this design. Only build after we have agreed on any modifications", random maps in 45 minutes, multiplayer over audio chirps borrowed from Chirp Pong and an online relay.

The skills solo builders wrote: route, handoff and a playtest loop

The custom skills in these runs solve two problems: a subscription that runs out and a context window that fills up. The RPG's developer wrote a "route" skill, in which Opus sends suitable work to Sonnet or Haiku and reviews the result, and a "handoff" skill that starts a fresh session with the same setup and only the context it needs.

The same developer added "Re-searcher", which checks research claims and stores them in Obsidian so later sessions do not work from "a summary of a summary", plus an art-direction skill and a JSON library of every image prompt that worked. One commenter's summary: "you only write the handoff skill after losing a session to context".

Godogen started in March with an orchestrator skill that planned, an executor that ran each piece in a forked context and a separate Gemini Flash agent that judged screenshots without seeing the code. By July its changelog shows the two skills merged, the Gemini pass dropped and the agent checking its own captured frames. The Battle Drones repository holds no skills; the dynamic workflows and agent-built tools did that job. In his words, "Tools are effectively free": Claude also built a model browser, a battle replay viewer and balance harnesses.

What it cost: subscriptions, not invoices

Every builder here paid a flat subscription, so none can price a single game. The one measured run time is Tree Appreciator's: about 3 hours 7 minutes and 154 tool calls from one prompt to a published page, by Claude's own reading of its session log.

The two compressed sizes are mine, measured with curl on 29 October; the rest are the builders' figures. On Stackness, as of 29 October 2026, Claude Code sits on 8 public stacks once bot accounts are removed, one of them mine, and three.js, Godot and Blender on none (how I count). The models sit with the other AI tools developers list.

Where the runs broke

The runs broke in game logic, in what the agent cannot see and on real hardware. Willison's Emberwick prompts read as a bug list: a tomb reached by dropping off a hill, companions that "vibrate back and forth near a wall", a quest reward that never fired because an NPC was missing, a save lost on refresh.

• The agent cannot see the screen. Godogen's author described floating assets and paths that lead nowhere; one Godot user called Claude Code with Godot "a total failure" for an RPG area because it could not tell where the doors were. Godogen's frame checks exist for this.
• Animation and model cleanup. Godogen's author wrote in March that "Animation is the main remaining gap"; its October README now advertises generated rigging and animation. A commenter on the RPG thread: "coding isn't the bottleneck it's the asset pipeline".
• Performance. Mecha Royale players report 15 fps on a GTX 1060 and memory past 1.4 GB. The agent tested Tree Appreciator only in software WebGL.
• Determinism. Math.sin and Math.cos differ slightly between Chrome and Safari, which broke Battle Drones replays until Claude wrote deterministic versions.
• People. Battle Drones' audio multiplayer gave Willison's wife "a nasty headache". He will not enter these games in game jams, and Bada Boom Taxi borrows from a film studio's IP.

Run it yourself

The run to copy is Battle Drones, public under Apache-2.0. Clone it, read web-session-prompts.md, which covers the sessions up to 13 October, and start a smaller game the same way: a design Q&A first, then the build, then one long "make it more fun" run with playtests.

git clone https://github.com/simonw/battledrones
cd battledrones
npm install
npm run dev

Then prompt your own game in claude.ai or Claude Code with "Build a mobile-friendly three.js game where...", and host it on GitHub Pages. If you want Godot instead, start from godogen's README, which sets up Claude Code or Codex for it. Before a long unattended run, write a handoff note or skill, so a full context window costs a restart rather than the session. Battle Drones itself was rewritten on purpose once, from Phaser to three.js at version 5, so expect at least one rewrite of your own.