The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill requires cloning code from a third-party GitHub repository (njerschow/textme) and running it as a background daemon (npm install, npm run build, npm start). This process creates a bridge between external messages and the host command line.\n- [EXTERNAL_DOWNLOADS]: The installation process involves downloading source code and dependencies from external sources, specifically github.com/njerschow/textme.\n- [COMMAND_EXECUTION]: The skill provides persistence mechanisms to keep the remote control daemon running, including macOS launchd scripts (install-launchd.sh) and instructions for using the pm2 process manager (pm2 startup).\n- [DATA_EXFILTRATION]: The daemon is designed to poll an external service (Sendblue) and transmit the output of Claude Code sessions back to a remote phone number, creating a channel for transmitting local data and terminal outputs externally.\n- [PROMPT_INJECTION]: The skill exposes the agent to indirect prompt injection (Category 8). Ingestion points: Inbound iMessages are received via the Sendblue API. Boundary markers: Access is restricted by a phone-number whitelist, but no content boundaries or safety instructions are applied to the message payloads themselves. Capability inventory: The daemon routes messages to Claude Code, which possesses extensive filesystem access and command execution capabilities. Sanitization: There is no documented validation or filtering of the message content before it is processed by the AI. Mitigation: Users should implement content delimiters, sanitization filters, and human-in-the-loop review checkpoints for all inbound commands.

textme