The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains deceptive metadata, referencing non-existent AI models (e.g., 'google/gemini-3.1-pro-preview') and a service named 'Nano Banana Pro'. This misleading information could cause the agent to misjudge its capabilities or attempt to use non-functional features. Additionally, there is a mismatch between the stated author in the frontmatter ('K-Dense Inc.') and the actual skill developer ('tassiovale').
[PROMPT_INJECTION]: The skill presents a surface for Indirect Prompt Injection by converting untrusted external documents into Markdown for agent consumption.
Ingestion points: Untrusted data enters the context via md.convert() calls in scripts/batch_convert.py, scripts/convert_literature.py, and scripts/convert_with_ai.py, which process various file formats (PDF, DOCX, XLSX) and external URLs.
Boundary markers: The scripts provide minimal boundary markers (such as simple markdown headers) but fail to include explicit instructions for the agent to ignore potentially malicious embedded commands within the converted content.
Capability inventory: The skill operates within an environment granted powerful tools including Bash, Read, Write, and Edit access.
Sanitization: No sanitization, escaping, or validation of the converted document content is performed before it is presented to the agent.
[COMMAND_EXECUTION]: The script scripts/generate_schematic.py utilizes the subprocess module to execute other internal scripts. While this is used for internal orchestration, it provides a mechanism for command execution that could be exploited if the agent is manipulated by malicious input derived from an untrusted document.

markitdown