The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to install two optional enhancement plugins from unverified GitHub users (Lum1104/Understand-Anything and mksglu/context-mode) using the platform's marketplace commands.
[INDIRECT_PROMPT_INJECTION]: The skill generates repository-wide instructions (AGENTS.md) by scanning all files in the project codebase, which could allow malicious code comments or documentation to influence downstream agent behavior.
Ingestion points: Scans all codebase files using tools such as find, grep, cat, and git as defined in SKILL.md steps 3 and 4.
Boundary markers: No explicit delimiters or instructions are used to separate codebase content from the agent's internal instruction generation logic.
Capability inventory: The skill writes and updates critical repository documentation (project-doc.md, AGENTS.md) and updates project state (state.json).
Sanitization: The skill lacks mechanisms to sanitize or filter codebase data before it is interpolated into agent instructions.

scan