The Agent Skills Directory

[PROMPT_INJECTION]: The skill operates as an orchestrator, receiving arbitrary user tasks and delegating them to various subagents. This architecture presents a surface for indirect prompt injection, where instructions embedded within a user's task or intermediate 'handoff' data could attempt to override the instructions of subagents.
Ingestion points: User task reception and classification in SKILL.md (Phase 0).
Boundary markers: While the skill uses a 'サブタスク入力テンプレート' (subtask input template) for delegation, it lacks explicit delimiters (like XML tags or triple quotes) and specific instructions to subagents to ignore potential commands embedded within the task data.
Capability inventory: The orchestrator delegates to agents with significant capabilities, including code implementation and editing (default agent), file system exploration (Explore agent), and security review (code-reviewer). It also directly executes commands like git status and ls.
Sanitization: The instructions do not define methods for sanitizing or validating user-provided task descriptions or the 'Handoff' information passed between execution steps.
[EXTERNAL_DOWNLOADS]: The README.md file contains instructions for downloading the skill from the author's repository using apm, degit, and the skills CLI (targeting github.com/bigdra50/skills). These are standard installation procedures from the vendor's own infrastructure.

orchestrator