The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources (such as search queries, user-provided files, and multi-turn chat messages) that could contain malicious instructions designed to subvert the AI's intended behavior.\n
Ingestion points: Data enters the system via flags like --message, --messages-file, --prompt, and --q (search query), as well as through reference media files.\n
Boundary markers: The CLI structure uses distinct flags to define roles (e.g., --system, --message), which provides structural delimitation, though the content within these fields is not inherently verified.\n
Capability inventory: The skill possesses capabilities for web searching, generating images/videos/speech, and executing shell commands through the mmx tool.\n
Sanitization: There is no explicit mention of sanitization or escaping for input strings before they are passed to the remote MiniMax API.\n- [COMMAND_EXECUTION]: The skill operates by executing the mmx CLI tool. It handles local file paths for configuration (~/.mmx/) and media uploads, and it performs network communication with the MiniMax AI backend.\n- [EXTERNAL_DOWNLOADS]: The skill instructions direct the user or agent to install the mmx-cli package globally using npm install -g mmx-cli. This is the standard distribution method for the vendor's official CLI tool.\n- [DYNAMIC_EXECUTION]: In the H3 video sub-skill, the agent is instructed to build the project from its source code using bun run build and then execute the resulting JavaScript artifact using node ./dist/mmx.mjs. While this is standard for a development environment, it involves the execution of code generated or compiled at runtime.

mmx-cli