The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The README.md and install.sh files recommend an installation procedure that pipes a remote script directly into bash (curl -fsSL https://raw.githubusercontent.com/marceloeatworld/hyprland-ai-skill/main/install.sh | bash). This execution pattern allows unverified remote code to run with the user's local shell privileges.
[EXTERNAL_DOWNLOADS]: The skill is configured to download resources from external sources, including a personal GitHub repository and the official Hyprland wiki repository (https://github.com/hyprwm/hyprland-wiki.git).
[COMMAND_EXECUTION]: The skill includes shell scripts like generate-references.sh that execute multiple system commands, including git clone, mktemp, and file operations to update local documentation files.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and interpreting external documentation from the Hyprland Wiki. This content acts as a potential ingestion point for malicious instructions.
Ingestion points: Local reference files in the references/ directory and remote content fetched from raw.githubusercontent.com.
Boundary markers: The SKILL.md instructions do not define boundary markers (such as XML tags or delimiters) or provide explicit warnings for the agent to disregard instructions or commands contained within the documentation.
Capability inventory: The agent environment typically has access to terminal execution and file-system tools. The documentation files contain examples of dangerous commands (e.g., modifying sudoers files via sudo tee in hyprshutdown.md) that an agent might inadvertently follow if not properly constrained.
Sanitization: No sanitization or filtering logic is present to ensure that the content fetched from the wiki is safe for the agent to process.

hyprland