Claude Fable 5 turned a four-year-old tweet into a playable 3D browser game in a single development session, and it did the heavy lifting without any human-chosen tech stack or design input.

Simon Willison fed the model the tweet’s brief description, a screenshot, and two supporting images. From that modest prompt Claude Fable 5 wrote the entire codebase, selected Three.js as the 3D engine, generated the game’s textures via an OpenAI model, and pushed each change to a GitHub repository. The repository was wired to GitHub Pages, so every commit appeared at a live URL within about half a minute. The result is a functional “Raccoon Heist” game that anyone can play in a browser today.

How the experiment unfolded

  • Minimal prompt – The original 2022 tweet listed a game concept and showed a single screenshot. Willison added two extra images for reference and a few lines of text. He gave no instructions about programming language, graphics library, or deployment.
  • Model-driven decisions – Claude Fable 5 chose Three.js, a JavaScript library for 3D graphics, as the rendering engine. It wrote the HTML, JavaScript, and asset-loading code needed to run the game.
  • AI-generated assets – An OpenAI model produced texture images, eliminating the need for a designer.
  • Continuous delivery – Each code push triggered GitHub Pages to rebuild the site, making the newest version visible in under 30 seconds. Willison watched the game evolve in real time, from a blank canvas to a fully playable prototype.

Why this matters

Traditional AI coding assistants have helped with autocomplete, bug-fix suggestions, or snippet generation when a developer supplies a detailed spec. Claude Fable 5 crossed a line: it accepted a high-level goal, chose the implementation stack, produced assets, and shipped a live product—all without further human direction.

Benchmarks such as SWE-bench measure how well models solve programming problems, but watching a model autonomously construct an end-to-end interactive experience shows a different capability. For rapid prototyping, the workflow proves that a developer can hand a model a concept and a handful of assets, point it at a deployment target, and receive a usable demo in minutes.

Who stands to gain, and who may be left behind

  • Start-ups and indie developers can iterate on game ideas or UI concepts without hiring a full engineering team.
  • Product teams may use the approach to generate proof-of-concept features for internal reviews.
  • Tool vendors that provide AI-driven code generation could see demand shift toward higher-level orchestration features—letting models pick frameworks and manage CI/CD pipelines.

Conversely, developers who specialize in architecture decisions, performance tuning, or art direction may find the most visible, low-complexity tasks increasingly automated. The model’s autonomy also raises questions about code quality, security, and licensing of AI-generated assets—issues that still require human oversight.

Limits of the current demo

  • Scope – The game is a modest 3D prototype; scaling the approach to large, data-intensive applications remains unproven.
  • Human scaffolding – Willison set up the GitHub Pages pipeline and monitored the model’s output. A completely hands-off system would need mechanisms to detect build failures or unsafe code.
  • Asset fidelity – AI-generated textures work for a simple game but may fall short for high-budget productions that demand photorealism or brand-specific art.

What to watch next

  • Tool integration – Expect tighter coupling between large language models and cloud build services, reducing the manual steps Willison performed.
  • Safety layers – Platforms may add automated code review or sandboxing to catch security flaws before deployment.
  • Benchmark evolution – New evaluation suites that test end-to-end product delivery could supplement existing code-completion metrics.

The takeaway is clear: AI can now move beyond “write this function” to “build and ship a complete, runnable product” when given a concise goal and some source material. The technology is still early, but the Claude Fable 5 demonstration hints at a future where the first draft of many software ideas is generated autonomously, freeing developers to focus on refinement, strategy, and creativity.