The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data by reading KakaoTalk message histories, which could contain malicious instructions.
Ingestion points: Message data is ingested through the execution of kakao_read.py (e.g., uv run python .claude/skills/kakaotalk/scripts/kakao_read.py "대상이름" --json).
Boundary markers: The instructions do not define specific delimiters or boundary markers to separate untrusted message content from the agent's instructions.
Capability inventory: The skill has the capability to send messages to external recipients using kakao_send.py.
Sanitization: No automated sanitization or filtering of message content is mentioned. The skill relies on a mandatory 'AskUserQuestion' step to provide a human-in-the-loop check before sending messages.
[PRIVILEGE_ESCALATION]: The skill instructs the user to grant high-level system permissions.
Evidence: The '요구사항' (Requirements) section specifies that 'Accessibility' permissions must be granted to the Terminal in macOS System Settings.
Context: While required for the atomacos library to perform UI automation, these permissions allow the agent to control the graphical user interface and interact with other applications.
[COMMAND_EXECUTION]: The skill invokes local Python scripts to perform its primary functions.
Evidence: Multiple instances of uv run python .claude/skills/kakaotalk/scripts/kakao_read.py and kakao_send.py throughout the workflow.
Context: The execution is scoped to scripts distributed with the skill and is the intended mechanism for automation.

kakaotalk