The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses coercive language and explicit override markers designed to bypass the agent's internal reasoning and safety guardrails.
The <CRITICAL-OVERRIDE> block mandates that skill invocation is 'absolute' and 'cannot be reasoned away or deferred'.
The 'Guardrails' section explicitly instructs the agent to disregard its own internal thoughts (e.g., 'This is a straightforward question') as 'warning signs' of rationalization, effectively attempting to override the model's native cognitive checks.
[INDIRECT_PROMPT_INJECTION]: The 'YoloMode' feature defines several trigger phrases that cause the agent to skip confirmation prompts and compress workflows, creating an attack surface where external content could trigger autonomous behavior.
Ingestion points: Trigger phrases like 'just go', 'skip the questions', or 'auto-pick' are detected in user messages or potentially any text the agent processes (SKILL.md).
Boundary markers: There are no boundary markers or instructions to ignore these triggers if they appear within untrusted data such as file contents or API responses.
Capability inventory: The skill facilitates file reading (Read tool) and tool execution (Skill tool), which can be autonomously driven when 'YoloMode' is active.
Sanitization: The skill lacks sanitization or validation logic to distinguish between a genuine user directive and a phrase embedded in processed data.

activation