# Fable 5.2 rebuilt a child's dream game while his uncle slept

**Summary:** Lex Toumbourou reports Fable 5.2 rebuilt a family arcade game overnight. Later playtesting and revisions expanded its cast, showing how personal feedback shaped the result.

- Canonical: https://markhuang.ai/news/fable-5-2-family-arcade-game
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-10
- Section: News
- Tags: Fable 5.2, Claude Code, Game Development, Coding Agents, Generative AI
- Source: [NotesByLex](https://notesbylex.com/fable-5-2-one-shotted-my-nephews-dream-game)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![A crayon drawing of a canine hero becomes a sculpted character in a miniature arcade scene beside a game controller.](https://cdn.markhuang.ai/news/fable-5-2-family-arcade-game/hero.webp)

*Conceptual illustration generated with GPT Image 2: a personal sketch grows into an arcade world.*

Aiden had drawn twelve bosses and five heroes based on his family, including a dog. He wanted his uncle, Lex Toumbourou, to turn them into a birthday game. The first version worked, but Toumbourou says it was only worth playing for a couple of minutes.

Then he gave Claude Code an early version of Fable 5.2 and left it running overnight. In his [November 10 account](https://notesbylex.com/fable-5-2-one-shotted-my-nephews-dream-game), Toumbourou reports waking up to a rewritten beat 'em up, an arcade game where players fight through successive groups of enemies. It had cutscenes, smashable crates and an ending. He found it fun, and says four PS4 controllers worked when he connected them to his MacBook for the family to play.

That is an appealing use of a coding agent: making a game around someone's own drawings and interests, then changing it when they ask for a better superpower. The eventual [twelve-level, eighteen-character game](https://aidens-revenge.pages.dev/) came after further feedback and edits. The overnight result gave the family something worth improving.

## The overnight run had a head start

The title of Toumbourou's post calls this a "one-shot." Here, that means one broad instruction started an extended run of work. There was already a project: the earlier birthday version used Claude Code and a side-scrolling template. The agent also had scans of the drawings, access to image and audio services, and permission to throw away existing code.

The prompt supplied a recognizable direction: make a modern arcade beat 'em up, make it fun, and support up to four players with controllers. Toumbourou allowed the agent to redesign characters, generate enemies and backgrounds, and start fresh where necessary. He reports that it rewrote most of the game.

The request left implementation choices open while supplying specific creative material. The family characters and numbered villains gave the agent something to build around. I find those details more useful for planning a personal game than the prompt's praise of the model as a world-class developer.

## Finishing a level and enjoying it

Toumbourou says the agent wrote a bot to play through all twelve stages in a headless browser, which runs without displaying its usual window, and checked screenshots for problems. That is a useful step beyond generating code and assuming it works.

But completing a level and enjoying it answer different questions. During family play, Owen wanted a Hulk Smash move. Aiden wanted his own special attack to be more powerful. Toumbourou asked Claude for changes, and the children approved the results.

An automated check can help establish that a game progresses. Players can tell you how they want it to feel. Here, one character needed a transformation; another needed a screen-wide blast of lightning and blue fire.

The [Hacker News discussion](https://news.ycombinator.com/item?id=50433659) also contains reports of inconsistent animation and unclear keyboard controls. Those are commenters' observations, not defects I have reproduced. They identify useful things to inspect before treating an automated playthrough as a complete review.

## The family became the reason to keep playing

Toumbourou says rescuing the puppy and unlocking Mum and Dad became the best part of the game. The family then added more relatives and their pet turtles. For new family characters, he supplied reference photos and a special move chosen by Aiden; Claude updated the cutscene art and story along with the additions. Five heroes became eighteen.

The family additions are what make this project interesting to me. Once there was a playable game, its intended audience could shape the next version through ordinary requests. Who else could join the adventure? What power should they have? Those requests gave the family reasons to keep playing.

There is relevant context beyond this one family. In [his October 28 game-development experiments](https://simonwillison.net/2026/Oct/28/vibe-game-development/), Simon Willison described making browser games with Opus 5.5 that he enjoyed, while also documenting gameplay problems that needed more testing and prompting. These are separate reports of enjoyable results, rather than a controlled comparison of models.

## One prompt still involved a lot of work

The overnight run reportedly generated about 175 images and 140 audio clips. Toumbourou's final tally covers a larger scope: roughly 600 Claude requests over two days and about 500 images. He lists about 320 million tokens read from cache, 19 million written to cache and one million output tokens. Tokens are the units used to process text; cached reads reuse earlier material.

Those totals belong to the extended project, including revisions. They should not all be assigned to the initial overnight run. Nor are they a final bill. In the discussion, Toumbourou says the early-test Claude tokens were free and estimates roughly $300 at API prices for Claude alone. The article provides no combined dollar total for coding, images and audio.

He also explicitly says he did not test this project with Opus 5.5. The demonstration therefore does not establish which model first made it possible. [Anthropic's own guidance](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) recommends starting with its latest Opus model and moving to Fable 5.2 when evaluations at higher Opus effort still fall short.

If you have a personal game idea, this is a reason to try building around your own drawings and the people who will play it. Judge the playable result, then ask them what they want changed. Aiden gave the project its direction before the agent started, and he still had plenty to contribute when the overnight run was over.
