The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes gh (GitHub CLI) commands via subprocess.run to interact with the GitHub API. These calls are well-contained, use capture_output=True, and do not pass unsanitized user input into a shell context.
[EXTERNAL_DOWNLOADS]: The skill requires the uv package manager and gh CLI to be installed on the host system. It does not perform any dynamic or unauthorized downloads during execution; it expects these tools to be pre-configured and authenticated by the user.
[DATA_EXFILTRATION]: No data is sent to non-whitelisted or untrusted domains. All network communication is performed through the gh CLI, which targets the user's configured GitHub instance (e.g., github.com).
[PROMPT_INJECTION]: The SKILL.md instructions include clear boundary markers and specific guidance on how the agent should handle feedback. It explicitly distinguishes between human review and bot feedback, and provides a structured workflow for addressing each without overriding safety guidelines.
[REMOTE_CODE_EXECUTION]: While the skill runs Python scripts using uv run, these are local scripts bundled with the skill. There is no pattern of fetching and executing code from remote or untrusted sources.

iterate-pr