The Agent Skills Directory

Command Execution via CLI Tools: The documentation utilizes the genkit CLI to manage development workflows, which involves executing shell commands provided by the user (e.g., genkit start -- <run_command>).
External Package Installation: The setup instructions guide users to install the genkit-cli globally and as a project dependency through package managers like npm.
Advanced Agent Capabilities: The skill includes examples of building agents with significant capabilities, such as shell access (runShell tool), file system operations (filesystem middleware), and sensitive actions (transferMoney tool). These features provide powerful functionality that should be reviewed for appropriate use cases.
Indirect Prompt Injection Surface: The agents described are designed to process untrusted data, such as user messages or external file contents, while maintaining access to tools like shell execution. This setup defines a surface for potential indirect instructions to be processed.
Ingestion points: The skill describes agents reading user input via chat.send and file contents through the read_file tool (e.g., in references/agents.md).
Capability inventory: Examples include subprocess execution via runShell and file system write access via the filesystem middleware (e.g., in references/agents.md).
Boundary markers: The provided agent examples do not explicitly implement delimiters or markers to isolate untrusted input within prompts.
Sanitization: The instructions do not specifically detail sanitization or validation routines for external content before it is interpolated into model prompts.

developing-genkit-js