The Agent Skills Directory

Metadata Poisoning (MEDIUM): The file references/usage-guide.md is inconsistent with the actual implementation in scripts/convert_pdf_to_images.py.
Evidence: The guide references a script named pdf_to_images.py and a command-line flag --html-output, neither of which exist in the provided skill files. This misleading metadata can cause an agent to attempt to execute non-existent commands or assume capabilities (HTML generation) that the skill does not possess.
Indirect Prompt Injection (LOW): The skill ingests untrusted external content (PDF files) and possesses file-write capabilities.
Ingestion points: fitz.open(pdf_path) in scripts/convert_pdf_to_images.py opens user-provided PDF files.
Boundary markers: None. The script does not use delimiters or instructions to ignore embedded content within the PDF structure.
Capability inventory: The script uses os.makedirs to create directories, pix.save to write image files, and zipfile.ZipFile to write archive files to the local filesystem.
Sanitization: No sanitization or validation of the PDF content or the output directory path (beyond basic existence checks) is performed.
Unverifiable Dependencies (LOW): The skill relies on an external library pymupdf.
Evidence: SKILL.md and scripts/convert_pdf_to_images.py require pymupdf>=1.23.0. While this is a common library, it is an external dependency that must be installed from a public registry.

pdf-to-image-preview