Building a web game with Claude Fable 5 and Playwright
According to Simon Willison, Claude Fable 5 built a playable web game from a tweet prompt, then used Playwright screenshots in Chromium to catch a mobile visibility bug before shipping static assets.
What happened
According to Simon Willison, he asked Claude Fable 5 running in Claude Code for web to build an entire raccoon-heist game from the contents of his 2022 tweet, and he links both a playable demo and a GitHub repo. He says the workflow began with an index page, a vendored copy of Three.js, and a generated textures script. Willison also describes procedural low-poly characters and a procedural WebAudio soundtrack with zero audio files. He reports that Fable smoke-tested the game in pre-installed Chromium with Playwright, captured desktop and mobile screenshots, noticed that the raccoon was invisible at mobile widths, and then fixed it. He says the deployed game uses static assets and makes no API calls.
Why it matters
Simon Willison describes a concrete agentic-development pattern: Claude Fable 5 built a playable web game, generated assets, and then used browser screenshots to catch a mobile visibility bug. That matters because it shows how a coding agent can move from generation to verification without leaving the browser, but only for this specific self-reported project.
Engineering impact
According to Simon Willison, the workflow combined vendored dependencies, generated textures, procedural audio, and Playwright smoke tests in Chromium. For engineering teams, the practical implication is to keep browser-executable checks in the loop when adopting code agents, especially for mobile layout regressions, while treating static assets as a deployment simplification rather than a performance proof.
Leadership impact
Simon Willison’s account suggests a management decision point: agentic tooling can compress prototyping, asset creation, and basic QA into one workflow, but the evidence is a single authored example, not a benchmark. Leaders should treat it as support for small, low-risk experiments where browser testing and asset boundaries are explicit, not as a general capability claim.
Editorial judgment
This is worth publishing because Simon Willison’s account shows a concrete agentic-dev workflow: code generation, asset generation, browser-based self-testing, and a mobile bug fix. The tradeoff is that it is a single-company-style self-report, so the value is in the reproducible workflow pattern, not independent proof of broader capability.
Watch next
The next useful check is whether similar agents still find cross-device UI issues when the task is less toy-like and the asset pipeline is less self-contained. Simon Willison’s post gives one narrow example, so the follow-up should compare outcomes on real product surfaces, with the same browser-based smoke-test loop and clearly defined mobile criteria.
Limitations & uncertainty
This analysis is limited to Simon Willison’s self-reported description of one project, so it supports workflow inference but not independent validation of Claude Fable 5’s general capability or benchmark performance.