The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands that interpolate user-provided input directly into a Node.js execution string (e.g., node ./scripts/resolve-url.js "<url>"). This pattern is susceptible to command injection if a user provides a URL containing shell metacharacters (e.g., ;, &, |) that are not properly sanitized before the command is invoked.
[PROMPT_INJECTION]: The skill processes data fetched from arbitrary external URLs to extract summaries, creating a risk for indirect prompt injection. A malicious research paper or webpage could contain hidden instructions designed to manipulate the agent's behavior during the summary generation or the interactive learning gate process.
Ingestion points: External content is retrieved via the fetch-fulltext.js script from user-supplied URLs.
Boundary markers: No specific delimiters or instructions to ignore embedded commands within the fetched content are provided in the extraction logic.
Capability inventory: The skill allows for local script execution via node, file system writes for the handoff document, and execution of the vidbyte CLI tool.
Sanitization: There are no explicit instructions or mechanisms defined to sanitize or validate the content retrieved from external sources before it is processed by the LLM.
[EXTERNAL_DOWNLOADS]: The skill references and suggests the installation of the vidbyte-skills Node.js package (npm install -g vidbyte-skills). This package is required for certain functionalities like the vidbyte retain command. While this appears to be the toolset provided by the author 'cerredz', users should verify the security and integrity of the package before global installation.

read-paper