The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill permits the autonomous installation of missing dependencies when operating in 'agentic' mode. Specifically, it instructs the agent to install packages if they are under 50KB without requiring user confirmation.
[COMMAND_EXECUTION]: In 'agentic' mode, the agent is directed to 'Resolve all blockers autonomously' and 'NEVER stop for human feedback'. This includes executing fix attempts (up to 3 retries) and installing software via the command line without human oversight or validation.
[INDIRECT_PROMPT_INJECTION]: The skill processes implementation plans from external files via read_file. It is a target for indirect prompt injection because:
Ingestion points: Reads the plan file specified by the user or the --plan parameter (SKILL.md).
Boundary markers: There are no delimiters or instructions to disregard potentially malicious commands embedded within the plan content.
Capability inventory: The skill has capabilities to read/write files, install dependencies, and execute arbitrary commands to resolve blockers.
Sanitization: The skill does not perform any validation or sanitization on the plan content, simply instructing the agent to 'Follow each step exactly'.

executing-plans