The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a multi-agent orchestration workflow that creates an inherent surface for indirect prompt injection, as findings and briefs are passed between different subagents.
Ingestion points: The skill ingests user-supplied tasks via [task] or $ARGUMENTS in SKILL.md.
Boundary markers: The instructions define strict prompt templates for subagents, including explicit "write scopes," "read-only boundaries" for research agents, and a mandatory "stopping rule" that requires subagents to report blocks rather than redesigning plans autonomously.
Capability inventory: The skill utilizes the Agent tool (subagent_type: "Explore" and "general-purpose") to orchestrate repository inspection, file editing, and validation. It also utilizes the $commit dependency for Git operations.
Sanitization: The orchestration logic requires post-wave reconciliation where all subagent results are verified against the manifest and the working tree, ensuring changes stay within assigned scopes.
[SAFE]: The skill demonstrates security best practices by including a configuration file (agents/openai.yaml) that explicitly disables implicit invocation (allow_implicit_invocation: false), ensuring the skill only executes when explicitly called by the user. The "Contract" section further reinforces a least-privilege approach to subagent orchestration.

claude-handoff