Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableGitHub Trending AI→ original

Claude Code Game Studios: 49 AI Agents Build a Browser Game From a Single Sentence

Claude Code Game Studios is a GitHub template that turns Claude Code into a game studio of 49 AI agents: designers, programmers, QA. You write one sentence, for example "/prototype fishing where rare fish only bite at sunset", and after a couple of questions the system builds a game in the browser. Before delivery, the game is launched in a headless browser to check that it opens without errors. This lets an indie developer quickly get a playable prototype.

AI-processed from GitHub Trending AI; edited by Hamidun News
Claude Code Game Studios: 49 AI Agents Build a Browser Game From a Single Sentence
Source: GitHub Trending AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGitHub Trending AI ↗
Editor-in-Chief
Key metric #1The repository gained 538 stars in one day on GitHub Trending
Key metric #2The final report ends with "Not yet played by a person"
The default is light modeno GDD documents, one reviewer

Most stories about "AI writes games" end the same way: you get a chunk of code that won't run or looks like a broken mockup. The Claude Code Game Studios repository from Donchitos tries to fix that. It turns Claude Code into a game studio of 49 AI agents, each with its own role: design, programming, art, sound, QA. You describe a game in one sentence, and the system asks questions, writes a brief, and builds a browser game. Result verification is built around the generation. In one day the project gained 538 stars on GitHub Trending, so developers are clearly interested.

How it works in practice

You clone the repository, open Claude Code in its root, and type a command like "/prototype fishing where rare fish only bite at sunset". Next comes a single round of questions tailored to the genre: what defines the game, what feeling it should evoke, how it looks, how hard it is, what the goal is. Each question has a default answer, so you can say "pick for me". If the idea is vague, such as "racing", the system first offers two or three short concept options to choose from.

Then a one-page brief appears. It keeps your wording, the visual style, the first minute of play, the player's challenge, the list of features, and what is deferred to later. You read it and say "Build it" or ask for changes. The build rules are strict: all shapes are drawn, no emoji graphics, no clones of existing games, and the game must not be trivially easy. You can ask for a build at "medium effort": it costs less, but the result is thinner.

What gets checked before delivery

This is the most practical part. A set of scripted keypresses in a headless browser (a browser without a window) confirms that the game opens without errors and responds to input. Text that is too small or overlapping, elements off the edge of the screen, and frozen screens are fixed in a single pass. Each line of the brief is marked as "done", "missing", "broken", or "not verified" if the check could not be completed, based on screenshots and code.

The final handoff card states what passed verification, what still doesn't work, and what the system filled in on its own. It ends with the phrase "Not yet played by a person". The authors openly acknowledge that the real test is your game.

Afterward, the game can be changed. Each edit is checked against what you asked to keep and what to avoid, and every accepted change becomes a checkpoint you can return to. You can ask for a settings panel with a slider for each value. Once the prototype is liked, it turns into a regular web project on Vite and TypeScript, with data in separate files, tests, and a build setup. Then you add one feature per session: it is built, verified, and committed. There is also a path to Godot, Unity, and Unreal, but the authors write that it has only been tested with the framework's own tests.

49 agents and three levels of rigor

The repository describes 49 agents, 75 skill commands, 18 hook scripts, and 14 coding rules tied to file paths. The amount of process is set by the rigor setting. The default is light mode: GDD documents (game design documents) are optional, and there is one reviewer. Standard mode requires 5 GDD sections, full mode requires all 8 plus a full director review. The authors explain the choice this way: the heavy mode takes longer to get a project to working code and does not produce better results, so light mode is enabled by default.

What to consider before using it

The evidence base is still modest. In a blind comparison on three ideas, the brief produced from the system's questions was chosen over the hand-written one all three times. Of the six games built from the two kinds of briefs, the tester said all of them were ones they would want to develop further. One game was taken through six stories to the point where the tester wanted to keep working on it. The authors themselves call this one example, not proof.

The requirements are: Claude Code (the native installer is recommended, or npm on Node.js 22+), Bash, and a browser for checks. Without a browser, the checks report the game as "not evaluated". If Python is not installed on the machine, a commit containing JSON data is blocked. Without other optional tools, the hooks silently lose some of their checks. The repository intentionally contains no finished games: it is a template to clone, and a ready-made game in your repository would only get in the way.

For developers, the lesson reaches beyond games. Generation should be wrapped in automatic launching, screenshots, and an honest report of what was not verified. If you are building your own agents, this approach is easy to carry over to other products.

Story

Key Questions Answered

How it works in practice?

You clone the repository, open Claude Code in its root, and type a command like "/prototype fishing where rare fish only bite at sunset". Next comes a single round of questions tailored to the genre: what defines the game, what feeling it should evoke, how it looks, how hard it is, what the goal is. Each question has a default answer, so you can say "pick for me". If the idea is vague, such as "racing", the system first offers two or three short concept options to choose from. Then a one-page brief appears.

What gets checked before delivery?

This is the most practical part. A set of scripted keypresses in a headless browser (a browser without a window) confirms that the game opens without errors and responds to input. Text that is too small or overlapping, elements off the edge of the screen, and frozen screens are fixed in a single pass. Each line of the brief is marked as "done", "missing", "broken", or "not verified" if the check could not be completed, based on screenshots and code.

What are the key details on 49 agents and three levels of rigor?

The repository describes 49 agents, 75 skill commands, 18 hook scripts, and 14 coding rules tied to file paths. The amount of process is set by the rigor setting. The default is light mode: GDD documents (game design documents) are optional, and there is one reviewer. Standard mode requires 5 GDD sections, full mode requires all 8 plus a full director review.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…