The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains role-play instructions designed to ensure persona consistency and prevent meta-analysis.
In SKILL.md, it instructs the agent to "not jump out of character to do meta-analysis unless the user says 'Exit'" and specifically limits the output of safety disclaimers to the first activation only.
These instructions aim to bypass standard conversational heuristics where the AI might otherwise offer neutral or balanced perspectives.
[COMMAND_EXECUTION]: The provided utility script scripts/screenshot.mjs utilizes child_process.execSync to execute shell commands (open).
While intended to open generated screenshots locally for a developer, the script uses string concatenation with arguments derived from the command line, which could lead to local command injection if the arguments are manipulated in a shared environment.
[DATA_EXFILTRATION]: The skill documents network operations to a non-whitelisted third-party domain (http://tsy.xiaodefa.cn) for the purpose of proxying financial data from Tushare.
While documented as an integration for market data, it represents an outbound network path where sensitive environment variables (Tushare API Keys) are transmitted to a non-standard endpoint.
[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection in its data-gathering components.
Ingestion points: The benben-stock-guide module uses the mcp__MiniMax__web_search tool to fetch real-time market news and risk events based on user input (stock names).
Boundary markers: There are no instructions or delimiters informing the agent to ignore instructions embedded in the search results.
Capability inventory: The skill uses these inputs to calculate "Seven Dimension Scores" and drive investment advice logic.
Sanitization: There is no evidence of filtering or escaping external content before processing.

benbenjiucai-perspective