The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute a locally resolved command (<global-refresh-command>) to synchronize changes to the skill library after modifications.
[DYNAMIC_EXECUTION]: Shell commands and file paths are dynamically resolved at runtime based on environment-specific bindings (e.g., skill_locations in SELF-IMPROVE.md), which are then used for file system operations and command execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from other agent skills during curation tasks, which presents an attack surface for instructions embedded in those skills to influence the curator's behavior.
Ingestion points: The agent reads and inventories skill files from sources such as <global-skill-source> and <repo-local-skill-source> as defined in SKILL.md and references/curation-policy.md.
Boundary markers: No explicit delimiters or isolation markers are defined for the content of the skills being processed to prevent the agent from mistaking data for instructions.
Capability inventory: The skill allows for significant file system modifications (patching, merging, archiving) and shell execution via the global refresh command.
Sanitization: The curation policy includes explicit instructions to identify and remove secrets, private data, or repo-specific context during the promotion process, which serves as a security best practice for data handling.

skill-library-curator