The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md direct the agent to execute a local shell script (scripts/setup-agents.sh) to initialize or update subagent tiers. This is a standard administrative task for configuring the local development environment.
[DYNAMIC_EXECUTION]: The setup-agents.sh script dynamically generates instruction files for subagents (~/.claude/agents/*.md) and uses an inline Python snippet to modify the tool's JSON configuration (~/.claude/settings.json). The code generated is based on hardcoded templates within the script.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from an external hook (classification labels) and user-supplied requests to form a "briefing" for subagents.
Ingestion points: SKILL.md reads Route: metadata from the system context and takes the user's request verbatim.
Boundary markers: The skill composes a structured briefing but does not explicitly use delimiters or "ignore" instructions for the subagent.
Capability inventory: The skill uses the Agent delegation tool to spawn subagents and shell execution via setup-agents.sh for filesystem writes.
Sanitization: No explicit sanitization or filtering of the user's input or hook metadata is described before it is passed to subagents.
[PERSISTENCE]: The skill modifies the user's local configuration directory (~/.claude/) to store agent definitions and permission settings (e.g., denying general-purpose) that remain in effect across different sessions.

dispatch