The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation provides instructions for executing various command-line utilities and local scripts for PDF processing.
Evidence: Calls to qpdf, ghostscript (gs), verapdf, and poppler-utils (pdftotext) are documented in references/legacy-utilities.md and references/batch-and-accessibility.md.
Evidence: The form-filling workflow in references/form-filling.md references a suite of local Python scripts (e.g., extract_form_field_info.py, fill_fillable_fields.py) for automated document manipulation.
[INDIRECT_PROMPT_INJECTION]: The skill uses LLMs and vision models to extract data from PDFs, which creates an inherent attack surface for instructions embedded within processed documents.
Ingestion points: The extractInvoice and extractComplexTable functions in references/ai-extraction-patterns.md ingest raw text and images extracted from PDFs.
Boundary markers: Not explicitly shown in the prompt snippets, though the instructions recommend directing the AI to specific high-value pages.
Capability inventory: The skill has capabilities to write files (fs.writeFile, doc.save) and execute subprocesses (python scripts/..., qpdf, gs) as part of its document processing pipelines.
Sanitization: The skill demonstrates a strong pattern of using Zod schemas to enforce structured output and validate data extracted by AI models, reducing the risk of accidental instruction following.

pdf-tools