The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a shell script intended for execution by the host agent to determine the appropriate authoring backend. This script invokes system commands such as which, vskill, and node, and can start a local evaluation server via vskill eval serve.
[INDIRECT_PROMPT_INJECTION]: The detection script in SKILL.md interpolates raw user input directly into a shell command: vskill skill new --prompt "<USER_PROMPT>". This pattern creates a command injection surface if the user input contains shell-active characters (e.g., backticks, subshell syntax, or unescaped quotes), as the skill does not provide explicit instructions for sanitization or escaping.
Ingestion points: The <USER_PROMPT> placeholder within the bash block in SKILL.md.
Boundary markers: The placeholder is wrapped in double quotes, which are insufficient to prevent all forms of shell-based injection.
Capability inventory: Execution of arbitrary CLI commands via the host agent's shell environment.
Sanitization: No sanitization, validation, or escaping logic is included in the skill's instructions.
[EXTERNAL_DOWNLOADS]: The skill documentation directs users to install external packages, such as vskill via npm. These represent external dependencies required for the skill's meta-authoring functionality.

skill-builder