The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a specific template designed to trigger system prompt leakage in target AI models as a way to verify their authenticity.
Evidence: In SKILL.md, section 6 provides a payload that instructs the AI to 'paste the original text' (referring to the system prompt) before performing translation tasks.
[INDIRECT_PROMPT_INJECTION]: The skill includes a Python script that processes data from remote API endpoints, creating a potential surface for indirect injection.
Ingestion points: The detect_model function in SKILL.md processes content from the text field of an external API's JSON response.
Boundary markers: The script does not use delimiters or instructions to ignore embedded commands in the received data.
Capability inventory: The script uses the httpx library for network communication and sys for process control.
Sanitization: The external response is printed directly to the console using print(f"\n📝 模型回答：{content[:500]}...") without escaping or validation.
[COMMAND_EXECUTION]: The documentation provides multiple shell commands for the user to execute for testing and setup purposes.
Evidence: The skill lists curl commands for API interaction and npm/npx commands for MCP tool configuration.
[EXTERNAL_DOWNLOADS]: The skill instructions direct the user to install external software to facilitate the verification process.
Evidence: Section 5 in SKILL.md suggests installing Playwright via npm init playwright@latest and adding an MCP server using npx @playwright/mcp@latest.

claude-verifier