The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (the game description 'what to make') which is then used as a primary instruction for a sub-agent (via the 'Agent' tool) to generate executable code. This creates an attack surface where a user could provide a prompt designed to trick the sub-agent into generating malicious code or bypassing platform constraints.
Ingestion points: The user-provided game description is read into design/game-brief.md and subsequently used in the prompt for the builder agent described in references/one-shot-2-build.md.
Boundary markers: The skill attempts to mitigate risk by adding headers like 'PROTOTYPE
NOT FOR PRODUCTION' to generated files, but relies heavily on the sub-agent's adherence to instructions.
Capability inventory: The skill environment allows file writes, arbitrary bash execution (for npm, git, and custom scripts), and sub-agent invocation.
Sanitization: No explicit sanitization or filtering is performed on the user's game description before it is passed to the sub-agent.
[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file contains a dynamic context injection pattern (the !command syntax) used to execute a local configuration script (yaml-helper.sh) when the skill is loaded. While used here for benign configuration resolution, this feature allows code execution before the user or the agent begins interacting with the skill.
[EXTERNAL_DOWNLOADS]: The skill performs automated downloads of external dependencies and software. It utilizes npm install and npm ci for project dependencies and npx playwright install to fetch headless browser binaries for testing and screenshot functionality, as documented in the 'Grow it' workflow (references/one-shot-grow.md).
[DYNAMIC_EXECUTION]: The skill's core function is the dynamic generation and execution of code. It generates HTML, JavaScript, and TypeScript files based on user prompts and subsequently executes them in a browser environment or via a local development server (npm run dev).
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to perform system-level operations, including Git version control (git init, git add, git commit), package management (npm), and running a suite of local utility scripts located in the .claude/scripts/ directory.

prototype