The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a multi-agent orchestration workflow that is susceptible to indirect prompt injection. Ingestion points: In Phase B and E, the skill ingests a design-sketch task and grounding artifacts (derived from the how and why skills), which may contain untrusted content from the codebase or user input. Boundary markers: The prompt template in references/runner-prompt.md lacks explicit boundary markers or delimiters for the interpolated variables, such as the task and grounding artifacts. Capability inventory: The skill delegates implementation tasks and orchestrates multiple model runners that can define and structure code. Sanitization: There is no evidence of sanitization or validation of the ingested artifacts before they are passed to the parallel runner models.

architect