The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingests untrusted data from user-provided files (PDF, DOCX, PPTX, XLSX, HTML, images, audio, ZIP, and YouTube URLs) through the markitdown tool. This content is then interpolated into the agent's context to classify the document and generate HTML output.
Ingestion points: The <file> argument provided to the /create-html command in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the workflow.
Capability inventory: The skill has the capability to read files, execute shell commands (uvx), and write files to the local filesystem.
Sanitization: No sanitization or filtering of the converted Markdown content is performed before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill executes shell commands using uvx markitdown <file>. This involves running an external utility on arbitrary user-supplied files. While markitdown is a known tool, processing complex file formats (like PDF or DOCX) from untrusted sources involves inherent risk if the underlying parser contains vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill relies on uvx to fetch and execute the markitdown package. markitdown is an official utility from Microsoft. The workflow uses this tool to handle various document formats before processing them into HTML.

create-html