The AI‑crafted game features five family heroes, twelve bosses, and unlockable relatives, all generated by Claude Fable 5.2.
*An AI model turned a handful of sketches into a multi‑level beat‑em‑up overnight. The result impressed a nine‑year‑old and rewrote dozens of lines of code in minutes.*
On a November morning in 2026, Lex Toumbourou opened his laptop to find a complete beat‑em‑up game waiting for him. The title, built from his nephew Aiden's hand‑drawn design docs, looked and sounded like a modern Ninja Turtles arcade classic. The surprise came from Claude Fable 5.2, an AI model the author had been invited to test on his "most ambitious use case." Within a single night the model turned sketches, a list of heroes, and a set of keys for OpenAI and ElevenLabs into a playable, four‑controller experience that delighted a nine‑year‑old on his birthday.
The author fed Claude Fable 5.2 a single prompt that listed the desired heroes, villains, graphics, sound, and controller support. The prompt invoked OpenAI's ChatGPT Images 2.5 for visuals and ElevenLabs for audio, supplying key file paths. Within hours the model produced roughly 175 images and 140 audio clips, rewrote existing code, and generated cutscenes, crates, and an ending. It also wrote a bot to play through all twelve stages in a headless browser, capturing screenshots for quality checks. When the author woke up, a fully functional game was ready to run on a MacBook with four PS4 controllers attached.
Fable 5.2 generated a full sprite set for the five family‑based heroes, including a dog named Doggo, and created spritesheets for twelve hand‑drawn villains. It added new henchmen sprites, background layers, and power‑up icons. The model also produced 175 image assets covering character art, cutscene frames, and level backgrounds, plus 140 audio clips for sound effects and voice lines. All assets were assembled into a side‑scrolling template originally sourced from AIOriented.dev, allowing the game to run on a standard controller setup.
After the initial playtest, the older nephew Owen requested a Hulk‑style smash. The author asked Claude to make Owen transform into a Hulk when he grabs his first power‑up, and the model delivered the change. Aiden demanded the most overpowered weapon; Claude responded with a screen‑wide blast of lightning and blue fire. The family later added uncles, aunties, grandma Ahma, grandpa Ah Gong, great‑grandma Tai Ma, and pet turtles Goldie and Billy as unlockable characters. For each, the author supplied reference photos and a special move, prompting Claude to redraw cutscene art and update the story accordingly.
The rapid turnaround—from raw sketches to a polished, four‑player arcade experience—shows that large language models can replace large development teams for certain genres. The ability to generate assets, write code, test gameplay, and iterate on design within a single night suggests a new workflow where creators act as prompt engineers rather than traditional programmers. If similar results scale to larger, more complex titles, the economics of indie game production could shift dramatically, pressuring conventional pipelines to adopt AI‑first pipelines.
The June version used Claude Code Opus 4.8 and Fable 5 with a side‑scrolling template from AIOriented.dev, but it was only fun for a couple of minutes.
Fable 5.2 generated roughly 175 images and 140 audio clips for the game.
The prompt instructed Claude to use OpenAI's ChatGPT Images 2.5 for images, ElevenLabs for sound effects, and provided file paths for both API keys.
Fable 5.2 proved that an AI can not only write code but also orchestrate art, sound, and testing to deliver a market‑ready game in hours. The experiment blurs the line between developer and prompt engineer, hinting at a future where indie studios lean on generative models to cut development cycles dramatically. As the technology matures, the industry will need to grapple with new standards for quality, ownership, and creative control.
Sources: [Hacker News – Fable 5.2 one-shotted my nephew's dream game]