The Agent Skills Directory

[SAFE]: The skill implements an indirect prompt injection surface because it processes untrusted task descriptions and codebase content to generate instructions for sub-agents. However, this is properly mitigated by the architecture.
Ingestion points: Task descriptions and codebase files (via Read, Grep, Glob, Bash) in SKILL.md.
Boundary markers: Instructions require sub-agent prompts to be structured with "Definition of Ready" and "Definition of Done" sections to delimit inputs and outputs.
Capability inventory: Agent (spawning agents), TeamCreate, TaskCreate, and restricted Bash (git commands).
Sanitization: The skill explicitly requires user confirmation via AskUserQuestion (Phase 3) before creating the team or spawning agents (Phase 4).
[SAFE]: The skill uses dynamic execution patterns by generating prompts for sub-agents at runtime. This behavior is expected for an orchestration skill and uses structured templates to maintain consistency.
[SAFE]: Tool access is well-scoped; for instance, the Bash tool is restricted to specific, non-destructive git operations.

team-stack