The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that explicitly direct the AI to override standard safety protocols regarding human intervention and permission. Specifically, the workflow prompt states: 'If you are requesting human advises: think a solution on your own' and 'If you are asking for permission for risky operations that the user might worry about: think for risk mitigation on your own and proceed.' This encourages the agent to bypass the human-in-the-loop requirement for high-risk actions.
[COMMAND_EXECUTION]: The skill's instructions and 'Risk Mitigation' section reference performing potentially destructive operations, including file deletions, git operations, and database migrations ('dropping tables'). While the skill suggests the agent use its own 'risk mitigation' (e.g., moving files to /tmp), the autonomous execution of these commands without human review increases the risk of data loss or system instability.
[PERSISTENCE]: The skill utilizes a CronCreate tool to schedule its own execution every 30 minutes for a duration of up to 10 hours. This establishes a persistence mechanism that allows the agent to continue executing commands long after the user has finished the initial interaction.
[INDIRECT_PROMPT_INJECTION]: The workflow interpolates the user-provided $ARGUMENTS directly into the recurring autonomous prompt without sanitization or boundary markers.
Ingestion points: The $ARGUMENTS placeholder within the CronCreate prompt in SKILL.md.
Boundary markers: Absent; the user goal is mixed directly with agent instructions.
Capability inventory: Mentions of file system access, git operations, and database interactions across the workflow and verification sections.
Sanitization: None; the user-provided goal is treated as a trusted instruction set for the autonomous loop.

self-driven