Simon Willison published four browser games on October 28, 2026, each built with Claude Opus 5.5, and none took more than four hours of his own work. The lesson worth stealing: start with one small prompt in a chat window, then keep asking for more, playtesting between every round.
The short answer
- What he built: Tree Appreciator, Emberwick, Battle Drones and Pumpkin Farmer.
- The effort: four hours or less of his own time per project, though Claude Code ran much longer on its own for one of them.
- The pattern: one chat prompt, staged additions, full playtests, then a coding agent once the idea works.
- The honest limit: his view that GPT-6 Astra builds games with less sense of fun is a personal impression, not a benchmark.
The four-step pattern
- Ask for a small, playable concept. His starter template is simple: a mobile-friendly three.js game, plus one sentence on what the game is.
- Add ambition in stages. Emberwick began as a village prototype, gained terrain and caves, and ended as version 11. Later rounds quadrupled the playable area and added mechanics and a plot.
- Play it start to finish and report what breaks. One bug he hit: a character the quest depends on ended up in the wrong place. A quick look would have missed it.
- Ask for fun out loud. The same request works as a standalone prompt, and he used it directly on Battle Drones.
Prompts that worked
None of these is clever. They are short, ambitious and specific about the feeling he wanted.
- Tree Appreciator: a spot-the-difference game between two generated trees, with a request to think hard about making it both fun and beautiful. It worked in one prompt, with no follow-up.
- Emberwick: a top-down RPG engine where walls turn transparent when you step inside a building. Follow-ups asked for basements joined by tunnels and a cave with a sleeping troll.
- Battle Drones: “Make it a LOT more fun,” then a request to port the 2D game to three.js.
- Pumpkin Farmer: an instruction to surprise and delight the player often, which produced many easter eggs.
When to leave the chat window
Chat was enough for the first two games. Pumpkin Farmer was different. He brainstormed a spec, saved it as a SPEC.md file, and gave it to Opus 5.5 in Claude Code, which worked for nearly two days. His advice: once a project outgrows a proof of concept, move to a coding agent such as Codex or Claude Code.
| Game | What it is | How he built it |
|---|---|---|
| Tree Appreciator | Mobile-friendly spot-the-difference game with two generated 3D trees; finished trees go into a “herbarium” | One prompt in the regular Claude chat, using three.js |
| Emberwick | Top-down fantasy RPG with a village, NPCs, basements, tunnels, a troll cave and a plot; about 15 minutes to finish | Regular Claude chat, from a first prototype (v01) to a final version (v11) |
| Battle Drones | Real-time strategy game for 2 to 4 players with seven maps, replays and rule-changing “mutators” | A July 2D version with Claude Fable 5, then Opus 5.5 prompted to make it more fun and port it to three.js |
| Pumpkin Farmer | Cosy isometric farming sim about growing and selling giant pumpkins | A brainstormed spec file, then Opus 5.5 in Claude Code for nearly two days |
What to check before you ship
- Play the whole thing. Quest-critical bugs hide in the middle of a game, not in the first minute.
- Read the code, or accept the risk. He practices “classic” vibe coding and does not read the code. That is fine for a toy, but nothing in his post checks safety or maintainability.
- Judge fun yourself. A game that runs is not a game that delights. He calls Pumpkin Farmer not his best and less fun than Battle Drones.
- Check rights for every generated asset. His post does not clear art, music or writing for commercial sale, so confirm the terms yourself.
- Expect pushback. Some players dislike AI-generated art, music and writing, and he would not enter these games in traditional game jams.
What he says about Opus 5.5 and GPT-6 Astra
He writes that Astra can build a game, but its sense of fun does not match Opus 5.5. Sending Emberwick’s starting prompts to ChatGPT with GPT-6 Pro produced Lantern Hollow, which he says looks great but plays far worse. He admits he has not run an exhaustive comparison, so read this as one builder’s anecdote.
Where the real bottleneck is now
He argues that building a game that plays well has become shockingly easy, and that the hard part is now coming up with a novel concept. Battle Drones shows what that opens up: he released its code under Apache 2.0, and a fork with a steampunk theme and new assets appeared the next day.
Questions and answers
How much of my own time should I expect to spend?
He reports four hours or less per project. Longer runs, like the near two days for Pumpkin Farmer, were Claude Code working on its own.
When should I move from chat to a coding agent?
When the idea works and the project has grown past a proof of concept. He suggests Codex or Claude Code at that point, and he fed Claude Code a SPEC.md file for Pumpkin Farmer.
Can I build and play on a phone?
Browser games run on mobile, and his starter prompt asks for a mobile-friendly game. He prompted Battle Drones from his phone.
Can I reuse the Battle Drones code?
The code is released under Apache 2.0, a permissive license. Read its terms before you copy anything.
Does this prove Opus 5.5 is the best model for games?
No. It is one author’s experience on a few projects, with no controlled comparison.
Terms in plain language
- Vibe coding: building software by prompting an AI without reading the code.
- three.js: a JavaScript library for 3D graphics in the browser.
- RTS: real-time strategy, a game style where you build units and fight over territory.
- Coding agent: an AI tool that edits and runs code on its own across many steps.
- Mutator: an option that changes the rules of a match, as in Battle Drones.
Sources
This article is based on his post; we did not run the games.
Facts read on 29 October 2026
