The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes PDF files from external sources, which introduces an indirect prompt injection surface. Maliciously crafted PDFs could contain text intended to influence the agent's behavior. However, the risk is significantly reduced by the structured workflow described in forms.md, which requires the agent to convert pages to images and perform visual analysis. \n
Ingestion points: PDF text and structure are parsed by pypdf and pdfplumber in scripts such as scripts/extract_form_structure.py and scripts/extract_form_field_info.py. \n
Boundary markers: The skill relies on process-based boundary markers (visual verification steps) rather than technical delimiters in the prompt content. \n
Capability inventory: The skill has the capability to write local files and execute external command-line tools. \n
Sanitization: No programmatic sanitization is applied to extracted text, but the multi-step verification process encourages safe handling. \n- [COMMAND_EXECUTION]: The skill documentation and instructions facilitate the use of external command-line tools and involve runtime library modification for specific functionality. \n
Evidence: SKILL.md and forms.md instruct the agent to utilize standard utilities like qpdf, pdftotext, pdfimages, and ImageMagick. These are standard tools for the intended task of PDF processing. \n
Evidence: The script scripts/fill_fillable_fields.py uses a monkeypatching technique to modify the pypdf library at runtime. The function monkeypatch_pydpf_method overrides DictionaryObject.get_inherited to resolve field retrieval issues in certain forms. This is a local, static code modification for improved functionality and does not expose the system to arbitrary code execution from untrusted input.

pdf