The Agent Skills Directory

Indirect Prompt Injection (LOW): The skill parses external codebase files to extract tool definitions and descriptions, which could contain malicious instructions designed to influence the behavior of the generated skill package.\n
Ingestion points: scripts/analyze_mcp.py reads .ts, .py, package.json, and pyproject.toml files from a user-provided directory.\n
Boundary markers: The skill does not use specific delimiters or safety instructions when incorporating extracted text into the generated SKILL.md templates.\n
Capability inventory: The agent is instructed to create new file structures and write scripts based on the parsed data, potentially propagating any injected content.\n
Sanitization: Extracted strings are handled as raw data and are not validated or sanitized before being placed into new files.

mcp-to-skill