The Agent Skills Directory

[DATA_EXFILTRATION]: The skill provides instructions for identifying and accessing local configuration files for AI coding tools (e.g., Claude Code, Cursor, Windsurf, Cline, IntelliJ). These paths (such as ~/.claude.json and mcp_settings.json) are sensitive as they contain tool configurations.
[PROMPT_INJECTION]: The skill's instructions involve reading content from local memory files and remote documentation using a broad set of capabilities (Bash, WebFetch, Read/Write), creating a surface for indirect prompt injection.
Ingestion points: SKILL.md (local memory files, protocol documentation URLs)
Boundary markers: Absent
Capability inventory: Read, Write, Edit, Bash, Glob, Grep, WebFetch as defined in SKILL.md frontmatter
Sanitization: Absent
[EXTERNAL_DOWNLOADS]: Fetches documentation and recommends using official packages from the @modelcontextprotocol organization, which is a well-known service for this protocol.
[SAFE]: The provided scripts and hooks (main.cjs, pre-execute.cjs, post-execute.cjs) are standard scaffolds or lifecycle handlers that do not contain malicious code.

mcp-builder