The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources (official library documentation) to update local skill files.
Ingestion points: Official documentation URLs (base_url) and content fetched by sub-agents such as reference-updater and reference-researcher.
Boundary markers: The skill implements a validation step for the mode parameter (check or apply) to prevent invalid values from being passed to sub-agents.
Capability inventory: The workflow has extensive capabilities including writing to the skills/ directory, updating CLAUDE.md, and initiating git commits through orchestration of sub-agents.
Sanitization: The skill provides explicit instructions to sub-agents (e.g., reference-researcher must not modify existing files) to prevent conflicting edits and maintain structure compliance.
[COMMAND_EXECUTION]: The skill orchestrates a suite of specialized agents that perform filesystem modifications and repository-level tasks. While this orchestration is standard for a maintenance tool, the autonomy given to sub-agents to modify project files based on external inputs represents an active capability surface.

update-skill