The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to perform multiple shell-based Git operations (git status, git commit, git push) and use CLI tools to retrieve system information (date) as part of its standard workflow.
[DATA_EXFILTRATION]: The instructions require the agent to automatically git push changes to a remote repository. In an autonomous execution mode, this behavior could be exploited to exfiltrate sensitive data if such information is committed to the repository during the 'Full Execution Mode'.
[PROMPT_INJECTION]:
Coercive Instructions: The skill uses aggressive language ('If you VERSCHLIMMBESSER this system, I will cut off your balls!') to influence agent behavior through role-played threats, which is a technique often used in adversarial prompting to override default constraints.
Indirect Injection Surface: The skill defines a 'Full Execution Mode' that processes user-supplied TODO lists with high autonomy ('NOW GET SHIT DONE!'). This creates a surface where instructions embedded in external data could lead to unauthorized actions.
Ingestion points: Processes user-provided task lists and planning requirements in SKILL.md.
Boundary markers: Absent. There are no delimiters or instructions to ignore embedded commands within the tasks being processed.
Capability inventory: Includes file system write access (docs/planning/) and shell command execution via Git in SKILL.md.
Sanitization: Absent. No validation or filtering is performed on the user-provided tasks before execution.

pareto-planning