The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes several shell commands to establish a recording environment, including docker, ffmpeg, xdotool, and socat.- scripts/record builds a Docker image and runs a container with host network sharing or port forwarding.- assets/docker/record.sh uses socat and /dev/tcp for network loopback functionality on macOS to ensure the container can reach the host's web application. This pattern, while sometimes flagged by security scanners, is used here for legitimate port forwarding.- assets/lib.mjs executes ffmpeg for screen capture and xdotool for simulating mouse and keyboard events.- [DYNAMIC_EXECUTION]: The skill dynamically generates and executes JavaScript code.- It creates a script.mjs file based on a template in SKILL.md and project-specific details.- assets/config.mjs uses dynamic import() to load a configuration file from a resolved file path (.demos/demo.config.mjs).- [INDIRECT_PROMPT_INJECTION]: The skill identifies and processes information from untrusted project files to drive its behavior.- Ingestion points: Reads files such as AGENTS.md, CLAUDE.md, and README from the user's project directory (identified in references/init.md).- Boundary markers: No specific markers are used to isolate instructions found in these project files from the agent's core instructions.- Capability inventory: The skill can build and run Docker containers, write to the .demos directory, and execute Playwright scripts that interact with web applications.- Sanitization: The agent is instructed to "work out" how to sign in and start the app from these docs, creating a risk that malicious instructions in the project docs could influence the generated script.

demo