The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external PDF documents to extract text, tables, and form metadata, which creates a surface for indirect prompt injection where malicious instructions embedded in a PDF could attempt to influence the agent's behavior.
Ingestion points: PDF content is ingested via pypdf, pdfplumber, and pdf2image across multiple scripts including extract_form_field_info.py and convert_pdf_to_images.py.
Boundary markers: The skill uses structured JSON formats (field_info.json, fields.json) to isolate extracted data, and the forms.md instructions include mandatory validation steps (check_bounding_boxes.py and visual image inspection) to verify output integrity.
Capability inventory: The agent has the capability to write PDF and image files and execute localized Python scripts to perform document manipulation.
Sanitization: There is no evidence of explicit sanitization of extracted PDF text before it is returned to the agent context.
[REMOTE_CODE_EXECUTION]: The script scripts/fill_fillable_fields.py performs dynamic execution by monkeypatching the pypdf.generic.DictionaryObject.get_inherited method at runtime. This is a targeted implementation intended to fix a specific bug in the library's handling of selection lists and does not execute arbitrary user input.
[COMMAND_EXECUTION]: The main skill instructions (SKILL.md) and advanced reference (reference.md) provide examples for using standard command-line PDF utilities including pdftotext, qpdf, pdftk, and pdftoppm for document manipulation.

pdf