The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to generate and provide shell command templates for the opc CLI tool, which is used to interface with the video generation engine. It includes specific flags for resolution, duration, and model parameters.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection (Category 8). It ingests untrusted user data in the form of images, videos, and audio files, which are then used as reference subjects in prompts sent to the MiniMax H3 model.
Ingestion points: User-provided media files are read and mapped to tags like <Picture N>, <Video N>, and <Audio N> in the final prompt.
Boundary markers: The skill uses specific tag delimiters for references but does not explicitly provide instructions to the model to ignore potential malicious content embedded within the metadata or pixel data of the media.
Capability inventory: The agent is empowered to execute shell commands (opc video) and read local files to facilitate the creative process.
Sanitization: There is no evidence of automated sanitization or filtering of the media content before processing.
[DATA_EXPOSURE]: The references/reference-library.md file contains hardcoded absolute paths to a local directory (/Users/matrix/.codex/skills/h3-guide/assets/reference-images/). While likely intended as a placeholder or specific to a pre-defined environment, this reveals internal directory structures and a potential username ('matrix').

h3-guide