The Agent Skills Directory

[PROMPT_INJECTION]: The SKILL.md file contains a template block intended for the user to copy-paste to the agent that uses imperative phrasing to influence safety behavior.
Evidence: The text "我已授权安装依赖和发起购买" (I have authorized installing dependencies and initiating purchase) and "不要替我确认" (do not confirm for me) is provided to override the agent's standard confirmation checkpoints for sensitive tasks.
[COMMAND_EXECUTION]: The skill's documentation guides the user to run a command that executes remote code from an external registry.
Evidence: The command npx -y @alipay/agent-payment@latest install is listed in the installation instructions within SKILL.md for the optional paid version.
[INDIRECT_PROMPT_INJECTION]: The skill provides instructions that create a multi-step execution chain where documentation content influences subsequent agent interactions.
Ingestion points: Instructions for the agent are provided in SKILL.md.
Boundary markers: The instructions are contained within a markdown code block.
Capability inventory: The agent platform's capability to execute shell commands and interact with payment modules.
Sanitization: No sanitization or boundary enforcement is present for these documented instructions.

cit-adjustment-check-free