The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill converts untrusted external files (such as PDFs, Word documents, and HTML) and YouTube transcripts into Markdown, which is then presented to the AI agent. This creates a vector for indirect prompt injection, as malicious instructions hidden within these source documents could influence or hijack the agent's behavior.\n
Ingestion points: Files and URLs processed by md.convert() in scripts/batch_convert.py, scripts/convert_literature.py, and scripts/convert_with_ai.py.\n
Boundary markers: The skill lacks explicit delimiters or warnings to the AI to ignore instructions found within the converted content.\n
Capability inventory: The skill has access to Bash, Read, Write, and Edit tools, and performs network operations.\n
Sanitization: There is no evidence of sanitization or filtering to remove potential prompt injection payloads from the converted text.\n- [DYNAMIC_EXECUTION]: The skill documentation and scripts facilitate the use of a plugin system that allows for the dynamic loading and execution of third-party code.\n
The MarkItDown instance can be initialized with enable_plugins=True, which loads external converters via Python entry points.\n
Documentation encourages users to find and install third-party plugins from GitHub using unverified hashtags.\n- [EXTERNAL_DOWNLOADS]: The skill references and downloads resources from external repositories and services.\n
Fetches the markitdown library from Microsoft's official GitHub repository.\n
Connects to OpenRouter (openrouter.ai) and Azure Document Intelligence for AI-powered processing.\n
Provides instructions for installing system dependencies like tesseract-ocr via standard package managers.\n- [METADATA_POISONING]: There is a discrepancy in the skill's attribution metadata.\n
The SKILL.md frontmatter lists the author as K-Dense Inc., while the provided author context for the skill is leonchaox. This mismatch in author identification could be deceptive regarding the skill's origin.

markitdown