The Agent Skills Directory

Subprocess Execution: The skill executes external binaries such as ffmpeg, ffprobe, and node to perform video and audio processing tasks. These are invoked using spawnSync with arguments passed as an array, which is a standard practice to mitigate command injection risks.
Evidence: Subprocess calls are found in scripts/assemble-index.mjs (for ffmpeg and ffprobe) and scripts/audio.mjs (for executing the audio engine).
Remote Resource Acquisition: The skill utilizes npx hyperframes to initialize projects, update skill definitions from GitHub, and fetch components from a remote registry. These operations involve downloading and potentially executing code from external sources to maintain the video generation environment.
Evidence: SKILL.md contains instructions for npx hyperframes skills update, npx hyperframes init, and npx hyperframes add to acquire remote resources.
Indirect Prompt Injection Surface: As a text-to-video generation tool, this skill ingests untrusted data from articles or notes to create video scripts and storyboards. This creates a potential surface where malicious instructions embedded in the source data could influence the agent's output.
Ingestion points: SKILL.md Step 1 processes user-provided text from BRIEF.md and visible-text.txt.
Boundary markers: The skill instructions emphasize using the input text as a source of information, but explicit adversarial boundary markers are not strictly enforced.
Capability inventory: The skill can execute shell commands via npx, perform file system operations, and run sub-agents.
Sanitization: Final assembly scripts include some basic escaping (e.g., in scripts/captions.mjs), but the initial narrative design phase is susceptible to instructional influence.
Autonomous Execution Mode: The skill features an "autonomous mode" that bypasses manual approval checkpoints for storyboards and visual designs. This reduces human oversight, which could allow unintended content to be processed through the video generation pipeline.
Evidence: SKILL.md Step 3 and Step 6 describe an autonomous mode that proceeds without user confirmation.
Verbatim Information Relay: The agent is instructed to relay the output of authentication status commands verbatim. Depending on the tool's behavior, this could inadvertently expose configuration details or login states in the session history.
Evidence: SKILL.md Step 0 instructs the agent to run npx hyperframes auth status and relay its output verbatim.

faceless-explainer