The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and process existing project documentation and instruction files which could theoretically contain malicious directives.
Ingestion points: The skill reads existing AGENTS.md files and project documentation in the target directory to learn project preferences.
Boundary markers: It uses explicit  and  blocks to delineate its managed content and preserve unrelated existing instructions.
Capability inventory: The skill is restricted to reading local files and writing the project's root AGENTS.md file. It does not perform network operations, install tools, or execute arbitrary shell commands.
Sanitization: The instructions explicitly direct the agent to omit secrets, temporary permissions, or absolute machine paths from the recorded result and to verify that unrelated content remains intact after edits.
[SAFE]: The skill implements strong safety principles, including a requirement for explicit user authorization before making changes, a 'disable-model-invocation' configuration, and clear boundaries for what actions are and are not authorized during setup.

setup