The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates and executes multiple transient scripts at runtime, including Python scripts for Mac Chrome fixes, PowerShell scripts for Windows Chrome fixes, and Bash scripts for plugin patching.
[REMOTE_CODE_EXECUTION]: The tutorial instructions include patterns for piped remote script execution, such as curl -fsSL https://bun.sh/install | bash, and direct execution of unverified third-party software installation commands via npx and uv.
[EXTERNAL_DOWNLOADS]: The skill installs several third-party packages from non-trusted repositories, including claude-mem from thedotmack and litellm from BerriAI, and downloads configuration guidelines from various unverified GitHub accounts.
[CREDENTIALS_UNSAFE]: The skill instructs users to provide sensitive API keys (e.g., Zhipu AI, DeepSeek) and automates writing these secrets into environment variables and shell profile files like ~/.zshrc.
[PERSISTENCE]: The skill creates a launchd persistence mechanism on macOS by generating and loading a Plist file (~/Library/LaunchAgents/com.claude-mem.litellm.plist) to ensure a background proxy service runs automatically on system startup.
[DYNAMIC_EXECUTION]: The skill uses sed to invasively patch the compiled source code of the claude-mem plugin (worker-service.cjs) to redirect its hardcoded API endpoints to a local proxy, which is a technique for traffic hijacking.
[PRIVILEGE_ESCALATION]: The instructions for the Chrome Gemini fix explicitly require the agent to guide the user into executing commands with administrative/root privileges on Windows and macOS.
[DATA_EXPOSURE]: The skill accesses and modifies sensitive application configuration files, including ~/.claude/settings.json and Chrome's Local State file, which contains browser-wide preferences and security flags.

laohan-jiaocheng