The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its core function of ingesting untrusted external data.
Ingestion points: The skill reads from the project codebase, research papers, notes, experiment logs, and existing documentation to ground its grilling questions.
Capability inventory: It has the capability to update the AGENTS.md file (which serves as a canonical instruction set for future agents), create new research documentation files, and initiate 'scout' or 'ship' tasks that involve project changes.
Boundary markers: There are no instructions to wrap ingested content in delimiters or to provide the model with 'ignore' directives for content found within external files.
Sanitization: The skill does not perform sanitization of the extracted data before synthesizing it into durable project records or agent instructions.
[COMMAND_EXECUTION]: The skill initiates external actions by dispatching tasks to other components of the agent system.
It dispatches 'scout' tasks for technical investigations and 'ship' tasks for delivering project modifications.
While the actual execution is delegated to 'crewmates' or platform 'machinery,' the skill constructs the briefs that drive these subsequent command-line operations.

grill-with-docs