The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted content from PDF files (via text extraction) and images generated from them (via visual analysis).
Ingestion points: forms.md (instructions for the agent to analyze PDF page images) and scripts/extract_form_structure.py (reading PDF layout and text).
Boundary markers: None identified in the instructions or scripts.
Capability inventory: The skill includes multiple scripts capable of writing to the file system, such as scripts/fill_pdf_form_with_annotations.py and scripts/fill_fillable_fields.py.
Sanitization: Extracted text data is used to drive form-filling logic without explicit sanitization steps.
[COMMAND_EXECUTION]: The skill instructions and scripts rely on various external command-line utilities for PDF processing and OCR tasks.
Evidence: SKILL.md and forms.md provide examples and instructions for using qpdf, pdftotext, pdfimages, pytesseract (which calls Tesseract), and magick (ImageMagick).
[REMOTE_CODE_EXECUTION]: A script performs runtime modification of a third-party library's internal logic to correct behavior for PDF form data extraction.
Evidence: scripts/fill_fillable_fields.py implements a monkeypatch on the pypdf.generic.DictionaryObject.get_inherited method.

pdf