The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill encourages the installation and use of external Python packages including fastmcp (maintained by PrefectHQ), the official mcp SDK, and django-mcp-server. It references official documentation and repositories from PrefectHQ and the Django Project, which are recognized as well-known and reputable sources within the development ecosystem.
[COMMAND_EXECUTION]: The documentation describes using the fastmcp command-line interface for routine administrative tasks such as listing available tools (fastmcp list), inspecting tool schemas (fastmcp inspect), and testing tools manually (fastmcp call). These are standard operational commands for developers using these tools.
[INDIRECT_PROMPT_INJECTION]: The skill documents an implementation surface that is susceptible to indirect prompt injection, as it involves an agent processing untrusted data to perform backend actions.
Ingestion points: Untrusted data enters the agent context through tool parameters, such as the reason field in the cancel_order tool or order_id in get_order_summary, as shown in references/tool-design.md.
Boundary markers: The skill recommends implementing clear boundaries using Python type annotations and Pydantic Field constraints to define the expected format and limits of tool inputs.
Capability inventory: Tools implemented using these patterns have access to the full Django environment via django.setup(), allowing read and write operations on the database ORM and potential interaction with background workers (e.g., Celery).
Sanitization: While the skill emphasizes rigorous validation and permission checks using Django services, it does not specifically instruct on sanitizing free-text inputs for potential LLM-directed instructions beyond standard data validation.

fastmcp-django