The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes plan files from the project directory and user-supplied arguments without sufficient isolation.
Ingestion points: Reads plan content from docs/plans/*.md and accepts input via $ARGUMENTS which are then interpolated into orchestrator and persona prompts.
Boundary markers: Lacks robust delimiters or instructions to ignore embedded commands within the {plan_content} and {project_context} variables.
Capability inventory: The skill can launch subagents and execute shell commands via the gemini CLI tool.
Sanitization: No input validation or sanitization is performed on the ingested content.
[COMMAND_EXECUTION]: The skill utilizes shell commands for both environment inspection and task execution.
Dynamic Context Injection: Uses the ! syntax in SKILL.md to run git, ls, head, and cat during skill loading to extract current project context.
External Tooling: Provides an option to execute the gemini CLI tool sequentially to process reviews for each persona.
[EXTERNAL_DOWNLOADS]: When the Gemini engine is selected, the skill transmits project context and plan content to external service endpoints via the gemini CLI tool.

team-briefing