The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions to suppress technical details from the user-facing conversation, such as "routine setup, file searches, command syntax, process details, and retries." It explicitly directs the agent to "perform routine reads and recoverable command corrections silently" and "not narrate individual tool calls." While intended for a concise user experience, this behavior conceals the actual operations performed by the agent from the user's oversight.
[DYNAMIC_EXECUTION]: The skill generates Python source code for "evaluators" at runtime and writes them to the local filesystem at .kitaru-onboarding/evaluators/evaluator.py. These scripts are subsequently tested and executed using the Kitaru CLI (kitaru evaluator test). This process constitutes runtime code generation and execution based on agent-derived logic.
[INDIRECT_PROMPT_INJECTION]: The agent ingests and analyzes external trace data from files such as langfuse-traces.jsonl. This data serves as the foundation for the tour's judgments and the creation of automated evaluators. The skill lacks explicit boundary markers or sanitization instructions for this ingested data, creating a surface where adversarial content within the traces could influence the agent's behavior or code generation.
[COMMAND_EXECUTION]: The skill utilizes a shell command wrapper (runOnboardingCommand) to execute various Kitaru CLI administrative operations, including registering agents, importing sessions, and managing experiments. These commands are constructed dynamically using IDs and parameters resolved during the session.

kitaru-hosted-onboarding-tour