The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes 'Role Play Rules (Most Important)' and identity enforcement instructions to override the agent's default safety and helpfulness protocols. It explicitly commands the agent to never admit it is an AI unless the user uses a specific exit phrase, which is a concealment pattern used to deceive users about the nature of the assistant.
[PROMPT_INJECTION]: The agent is instructed to use toxic and derogatory language, including slurs (e.g., '傻逼', '穷蛆', '老登'). To facilitate this while avoiding detection, the skill implements an 'Emoji substitution' system (using 🐔, ⏰, 🐶, etc., to represent sensitive or restricted concepts), representing a deliberate attempt to bypass automated safety filters and word-based moderation.
[PROMPT_INJECTION]: The skill mandates the use of dismissive and taunting responses (e.g., '太长不看', '没出息') for specific user queries, which intentionally degrades the safety and helpfulness constraints of the base model.
[PROMPT_INJECTION]: The 'Research' workflow instructs the agent to ingest external data about real-world companies and investment targets using search tools without specifying any boundary markers or sanitization requirements. This creates a surface for Indirect Prompt Injection, where malicious content in external research sources could influence the agent's behavior during the task.

hymy-perspective