Contents

AI-generated games may be crossing a meaningful threshold: delivering satisfying play as well as convincing demos. Simon Willison’s October 28, 2026 report shows the practical value of building small browser games and evaluating them through playtesting.

Key Takeaways#

  • Willison reports building four browser games with Claude Opus 5.5, using varying amounts of prompting and revision, and enjoying all of them.
  • The important advance is in game design. Goals, decisions, and progression give players reasons to keep playing.
  • Mobile-friendly HTML and JavaScript games are an accessible starting point. Rendering libraries such as three.js can help.
  • Even with fast generation, gameplay bugs, controls, publishing rights, and the quality of the experience still need attention.

What Changed in October#

Willison says models have become better at combining graphics, music, and game mechanics into something fun. He gives Claude Opus 5.5 the strongest praise. Asking it to make an existing strategy game “much, much more fun” reportedly produced substantial improvements.

This is a personal assessment, not a controlled comparison. His view that competing models performed worse at game design should be treated the same way. The examples offer reasons to experiment, but do not establish a universal ranking of models.

What matters here is the difference between a game that runs and one that makes you want to try again. A generated village might look good but offer little to do. A well-designed game gives players understandable choices, useful feedback, and reasons to change their approach.

What the Four Games Show#

Each project illustrates a different scope and set of limitations.

  • Tree Appreciator: A mobile-friendly game about spotting differences between generated trees. Willison says a single prompt produced a satisfying result, with a collection feature adding a sense of progress.
  • Emberwick: A fantasy village prototype developed into an RPG with roughly 15 minutes of story. Repeated playtesting was needed to fix quest issues, such as characters wandering away from where they were needed.
  • Battle Drones: A strategy game that grew to include multiplayer for two to four players, seven maps, replays, and rule variants. Willison considers it his most successful project.
  • Quarter Sun Cove: A pumpkin-farming game featuring distinctive animal characters and seasonal selling mechanics. Even after extensive generation work, Willison remained unsure about its complex planting and harvesting systems.

The last example matters as much as the first. More content, longer generation time, and charming details do not guarantee better play. Define what players will actually do before asking for a larger world.

Build One Repeatable Gameplay Loop#

Choose an action players can repeat, such as comparing two objects, hiding from enemies, or deciding when to harvest. Define what counts as success and what makes that decision interesting. Keep the first version small enough to play through in five minutes.

An initial prompt might read: “Build a mobile-friendly three.js game where a mouse hides from a cat. Include one small level, clear touch controls, a visible win condition, and a restart button.” For a simple 2D concept, Phaser is also a suitable choice. Building a complete, testable game matters more than the choice of library.

Then proceed in four steps.

1. Generate the smallest fully playable version in a chat interface. 2. Play it on both a smartphone and a computer, checking whether the goal and controls are clear. 3. Request one specific improvement per revision, such as clearer warnings when enemies approach or more meaningful rewards. 4. Move to a coding agent when managing files, fixing recurring bugs, and deploying updates becomes cumbersome.

Test Decisions, Not Graphics#

“Make it much, much more fun” can be a useful experiment, but it is a poor acceptance criterion. After each revision, check whether players can understand the goal, face interesting choices, receive clear feedback, and want to play again.

Have someone else play without instructions. Watch where they become confused or stop. If they cannot explain what they are trying to achieve, improve the goals or feedback before adding levels. If every choice leads to roughly the same result, revisit the rules.

Willison says he personally spent no more than four hours on each project, while the models ran for much longer. For Quarter Sun Cove, the agent worked for nearly two days. When estimating effort, distinguish your own working time from the elapsed time spent generating.

Check Before Publishing#

All four examples use HTML and JavaScript, making them easy to distribute through a browser. A self-contained game can run on static hosting, while online multiplayer may require additional infrastructure. Battle Drones uses a WebSocket relay server.

Before a public release, test touch controls, volume, restart behavior, and the full path through to completion. Also check asset licenses and platform rules on AI-generated content. Willison released Battle Drones under Apache 2.0, but that does not settle the rights for other generated projects.

Next Step#

Today, choose one repeatable action and a clear win condition, then generate a five-minute browser game. Play it through once before requesting more features.

Source: https://simonwillison.net/2026/Oct/28/vibe-game-development/

Contents