Claude Fable 5.2 reportedly built a complete 12-level beat 'em up from a child's drawings

A birthday ask turns into a game project

Claude Fable 5.2 reportedly produced a complete, playable 12-level beat 'em up in a single overnight run, working from a child's drawings, a folder of scans and one open-ended prompt. The account comes from a developer who writes as Lex on his blog, Notes by Lex, in a post that reached the front page of Hacker News.

The project began when Lex's young nephew Aiden asked for a custom beat 'em up for his birthday, a genre the family had been playing together on an old PS4. Aiden did not come empty-handed: his design documents specified five playable heroes based on family members (including the dog, Doggo), twelve numbered bosses, a loose story about rescuing the family puppy, and parents as unlockable characters.

Around January 2026, Lex considered a from-scratch indie game an unreasonable birthday ask and said so, but treated it as a chance to stress-test his agentic tools. A first version shipped for the June birthday, built mostly with Claude Code on Opus 4.8 and Fable 5 on top of a side-scrolling template from AIOriented.dev. It ran, but only held anyone's attention for a few minutes. Aiden kept asking about the missing pieces every time they met.

The overnight run

With early access to Claude Fable 5.2 and a nudge to pick his most ambitious use case, Lex pointed the agent at the unfinished game just before bed. The prompt gave it wide latitude: finish the game, aim for the feel of classic TMNT and Streets of Rage arcade brawlers, generate images through OpenAI's ChatGPT Images 2.5 and audio through ElevenLabs, use the scans of Aiden's drawings for the heroes and villains, and exercise its own judgment. The prompt also borrowed a motivational framing technique from a Latent Space Engineering article, which Lex suspects may no longer matter.

By morning the game was finished and, by Lex's account, genuinely fun. The agent had discarded most of the existing code and rewritten the game, generated roughly 175 images and 140 audio clips, and added unrequested touches including cutscenes, smashable crates and a proper ending. Reading the reasoning traces afterward, Lex found that the model had written a bot to play its own game, run all twelve stages in a headless browser, and inspected the screenshots for problems.

Playtests and rewrites

The first family session ran on four PS4 controllers paired to a MacBook, and produced very specific feedback. Older nephew Owen wanted a Hulk-style smash, resolved as a transformation triggered by grabbing his first power-up. Aiden, the game's namesake, insisted his special weapon be the strongest in the game and received a screen-wide lightning attack. Unlockable family members turned out to be the most compelling mechanic, so the roster grew from five heroes to eighteen — uncles, aunts, grandparents and the pet turtles — each added from a few reference photos and a chosen special move, with cutscene art redrawn and the story revised each time. The final rescue is the children's great-grandmother, who had died earlier in the year, drawn as an angel at Aiden's suggestion.

The final tally

Across two days, the project used roughly 600 Claude requests: about 320 million tokens read from cache, 19 million written to cache and 1 million of output, which Lex rounds to 340 million tokens in total. Asset generation came to about 500 OpenAI images, including sprite sheets, character designs, backgrounds and key art, plus 11 music tracks, roughly 100 sound effects and 90 announcer voice lines from ElevenLabs. The result is a 12-level, 18-character, four-player brawler with original art, music and an announcer.

Lex adds one caveat: he never ran the same task on Opus 5.5, so he cannot say whether the one-shot result is unique to Fable 5.2 or was already within reach of the previous model generation.

Why it matters

A finished multi-level game is a demanding target for agentic coding: it requires long-horizon coherence across gameplay code, art direction, audio and game feel, plus working integrations with external generation services. Even a single anecdotal success at that scale says something about where these systems now sit.

The most telling detail may be the self-verification. Rather than stopping at generated code, the agent built a QA harness, played its own game end to end and checked the screenshots before calling the job done. Verification, not raw generation, is widely viewed as the bottleneck for agentic software work, and this is a concrete example of it closing.

The story also sketches a new shape for personal software: a bespoke game that would once have meant months of hobbyist development, delivered as a birthday present with a child acting as creative director. The usual caveats apply — one enthusiast's report, no baseline comparison, and reliance on paid image and voice APIs — but as a data point on capability ceilings, it is a striking one.