The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses forceful, imperative language (e.g., "禁止" / "prohibited", "全タスクに適用" / "apply to all tasks") to fundamentally alter the agent's standard behavior. It explicitly forbids the agent from presenting options or asking for clarification, requiring it to "think as a genius" and act autonomously.
[INDIRECT_PROMPT_INJECTION]: The mandate for the agent to "decide and execute" without waiting for user approval increases the vulnerability surface for indirect attacks. If the agent processes malicious data (e.g., from a file or web search) while in this high-autonomy mode, it may be more likely to act on embedded instructions without human review.
Ingestion points: The skill triggers on any user instruction and processes data encountered during task execution.
Boundary markers: No specific boundary markers or delimiters are instructed for external data processing.
Capability inventory: While no tools are defined within the skill, it governs the agent's use of all available system capabilities (file writing, command execution, etc.).
Sanitization: The skill relies on the agent's internal safety filters and a specific "Exceptions" section that mandates confirmation for destructive or financial actions.
[SAFE]: The skill includes clear safety constraints that require human confirmation for irreversible operations (such as git reset --hard or file deletion), financial transactions, and public data sharing, mitigating the risks associated with increased autonomy.

autonomous-genius-execution