The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to clone a software repository from an untrusted source: https://github.com/leigest519/OpenGame.git. This repository is not associated with a known trusted organization or service.
[REMOTE_CODE_EXECUTION]: The installation process requires running npm install and subsequent npm run commands on the cloned repository. Since the repository content is unverifiable, this constitutes the execution of untrusted code on the host system.
[DYNAMIC_EXECUTION]: The framework generates web game code (HTML, JavaScript, and CSS) dynamically from natural language prompts and executes it using a headless browser (Chromium via Playwright) for automated evaluation. Running dynamically generated code, especially code that interacts with a browser environment, presents a significant security risk if the generation process is compromised.
[INDIRECT_PROMPT_INJECTION]: The framework is designed to ingest natural language prompts to drive the agentic code generation and evaluation process. This creates a vulnerability surface where a malicious prompt could inject instructions that influence the generated code's behavior or the evaluation pipeline.
Ingestion points: User-provided game design prompts passed to the generate command.
Boundary markers: None explicitly documented in the skill instructions to separate prompt instructions from the execution context.
Capability inventory: The skill has the capability to write files to the local disk, execute shell commands via npm, and perform network requests to LLM and VLM endpoints.
Sanitization: There is no evidence of sanitization or safety filtering for the generated web code before it is executed in the headless browser.

opengame-agentic-game-creation