The Agent Skills Directory

[PROMPT_INJECTION]: The documentation in references/security.md contains instructional examples of prompt injection phrases (e.g., "ignore previous instructions") to educate users on how to recognize and mitigate such threats in converted document content; these are descriptions of risk, not malicious instructions.\n- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted external data by converting heterogeneous document formats into Markdown for LLM use.\n
Ingestion points: Document files are ingested through methods like convert_local() and convert_stream() in SKILL.md and implemented in scripts/batch_convert.py.\n
Boundary markers: Documentation in references/security.md recommends explicit labeling of output as untrusted and using semantic boundaries.\n
Capability inventory: The skill can perform network requests and call external Azure AI services if configured with credentials as documented in references/cloud_and_ocr.md.\n
Sanitization: The skill uses defusedxml for XML parsing and recommends sanitizing links in output text as described in references/security.md.\n- [EXTERNAL_DOWNLOADS]: The skill references and downloads configuration or code from trusted repositories and well-known services.\n
Resources: Fetches metadata from the arXiv API (arxiv.org), uses Azure AI services (azure.com), and installs Microsoft's official markitdown packages from GitHub repositories.\n
Context: These external operations are optional, require user-supplied credentials, and are documented as external data-handling paths in references/security.md.\n- [DYNAMIC_EXECUTION]: The skill provides an opt-in plugin architecture that loads code from package entry points.\n
Pattern: MarkItDown(enable_plugins=True) in SKILL.md and scripts/batch_convert.py enables this functionality.\n
Safety: This functionality is disabled by default, and the references/mcp_and_plugins.md file provides a comprehensive trust checklist for auditing plugins before activation.

markitdown