The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection by interpolating untrusted user input ($ARGUMENTS) directly into the prompts for various subagents such as the business analyst, backend architect, and deployment engineer. \n- Ingestion points: The $ARGUMENTS variable, representing external user input, is used throughout SKILL.md across all development phases (Discovery, Implementation, Testing, and Deployment). \n- Boundary markers: There are no delimiters, XML tags, or "ignore embedded instructions" warnings surrounding the interpolated $ARGUMENTS variable to distinguish it from the agent's instructions. \n- Capability inventory: The subagents targeted by these prompts possess high-privilege capabilities, including implementing backend logic (backend-architect), configuring CI/CD pipelines (deployment-engineer), and accessing performance monitoring (performance-engineer). \n- Sanitization: The skill does not implement any escaping, validation, or filtering of the user-provided arguments before they are passed to the subagent logic.

backend-development-feature-development