The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from a wide variety of file formats (PDF, images, audio, and YouTube transcripts) and converts it to Markdown for agent consumption. This content is then presented to the agent, creating a surface where malicious instructions embedded in these files could influence agent behavior.\n
Ingestion points: The scripts scripts/batch_convert.py, scripts/convert_literature.py, and scripts/convert_with_ai.py call the markitdown library to ingest and convert external file content.\n
Boundary markers: While the scripts add metadata headers to the output, they do not include explicit instructions or delimiters for the agent to treat the converted content strictly as data or to ignore any embedded commands.\n
Capability inventory: The skill is granted significant permissions including Bash, Read, Write, and Edit tools, and the provided Python scripts perform various file system operations.\n
Sanitization: The skill performs direct conversion without sanitizing or filtering the resulting text for potential instructional content before it is processed by the agent.\n- [PROMPT_INJECTION]: Instructions in SKILL.md under the 'Visual Enhancement' section attempt to override default agent behavior by mandating that scientific schematics 'should be generated by default' and directing the agent to use a specific external skill (scientific-schematics) for all new documents, regardless of the user's explicit request.

markitdown