Claude One-Button Game Creation

This project generates batches of compact, browser-playable one-button games with crisp-game-lib. Agents turn random tag combinations into ten independent game designs and implementations, verify that each game runs in a browser, and present the batch without ranking it.

See GAMES.md for published examples and screenshots. Published builds under docs/ are curated separately and may be changed only by explicit human instruction.

Setup

Prerequisites:

• Node.js 18 or newer with npm;
• Git and Bash;
• network access for npm, the agent-skill repository, Playwright's Chromium download, and the games' CDN-hosted crisp-game-lib scripts.

npm ci
npm run skills:install
npm run playwright:install
npm test

The skills command copies this project's three required skills from abagames/agentic-gamedev-skills into .agents/skills/. It replaces only those three skill directories and preserves other content under .agents/. That directory is local setup output and is not tracked. npm test verifies that the required skill files exist and that the smoke-test runner can launch Chromium. To intentionally use a different upstream revision, set AGENT_SKILLS_REF to its commit SHA when running npm run skills:install.

On Linux, if Chromium reports missing system libraries, install Playwright's OS dependencies with the privileges appropriate for your environment:

npx playwright install-deps chromium

No batch state needs to be initialized for generation.

Generate a batch

The normative one-game procedure is workflows/game-generation.md. Read it in full, then execute it ten times. This repository does not provide an npm command that creates games by itself; the workflow is executed by a coding agent that can read and modify the repository. From the project root, ask it:

Follow AGENTS.md and generate and present one batch of ten games.

Before starting, confirm that tmp/games/ is absent or empty. If it contains anything, stop and ask the user what to do. Do not delete, replace, move, or archive existing games without user direction. Starting empty makes every directory in the completed batch unambiguously new.

For each game:

1. Start a fresh generation context.
2. Select three fresh tags:

bash node scripts/random_tag_selector.js -n 3 -f text

1. Design and implement one game according to the workflow.
2. Write its three deliverables under tmp/games/<slug>/.

Repeat until the batch contains exactly ten new games. Each context works only on its own game; it does not read or rank the rest of the batch.

Verify the batch

Verify the directory count, required deliverables, and real-browser runtime of all ten games:

npm run batch:check

The command exits unsuccessfully if the batch does not contain exactly ten game directories, if a required file is missing, or if any smoke test fails. Repair every failure and run the command again. Smoke testing verifies that the games can be presented; it does not measure quality or compare games.

Present and stop

Present an unranked list of the ten slugs and their README.md files. The standard workflow ends there.

The following require a separate, explicit request and are not automatic continuation steps:

• comparative evaluation or defect annotation;
• human ranking, comments, selection, or winner records;
• tuning or Polish;
• retrospective analysis;
• copying or publishing a game under docs/.

Output contract

tmp/games/<slug>/
├── index.html    # Browser entry point
├── main.js       # Game implementation
└── README.md     # Six-section design description

The batch is complete when tmp/games/ contains ten new game directories and all ten browser smoke tests pass.

Repository layout

Archived workflow

The former generate→evaluate→human-select→Polish process is preserved under archive/retired-batch-selection/ for historical reference. It is not part of the current generation flow and must not be started automatically after a batch is presented.