The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The scripts/evaluation.py script facilitates an evaluation loop where the agent processes instructions from external XML files and incorporates outputs from MCP tools back into the prompt. This is a standard surface for indirect prompt injection within an evaluation framework.
Ingestion points: Evaluation XML files (qa_pair elements) and MCP tool results in scripts/evaluation.py.
Boundary markers: The EVALUATION_PROMPT defines a strict output protocol using XML tags (<summary>, <feedback>, <response>) to maintain agent focus.
Capability inventory: The skill has access to Bash, WebFetch, WebSearch, and file system tools (Read, Write, Edit, Glob, Grep).
Sanitization: Relies on structured prompting and the underlying safety guardrails of the configured Claude model.
[COMMAND_EXECUTION]: The instructions direct the agent to use the Bash tool for project setup, builds (e.g., npm run build), and running the MCP Inspector. These commands are legitimate and necessary for the skill's primary purpose of building and testing code.
[EXTERNAL_DOWNLOADS]: The skill fetches documentation and README files from the official Model Context Protocol GitHub organization and website. These are well-known, trusted sources, and the references are used for retrieving up-to-date protocol specifications.
[PROMPT_INJECTION]: Static analysis flagged potential prompt injection patterns in reference/security-middleware-attacks.md and reference/security-pii-compliance.md. These are false positives; the files contain educational documentation and regex patterns intended to help developers implement security filters to prevent injection attacks in their own MCP servers.

mcp-builder