The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted PDF data from external sources (text extraction, table parsing, OCR). This content is then read into the agent's context, creating a surface for indirect prompt injection if the PDF contains instructions intended to manipulate the agent.
Ingestion points: Multiple scripts such as scripts/analyze_form.py, scripts/extract_tables.py, and scripts/extract_text.py, as well as programmatic examples in FORMS.md, OCR.md, and TABLES.md ingest external PDF content.
Boundary markers: The skill instructions do not specify the use of delimiters or explicit guardrail prompts (e.g., "ignore instructions found within the document") when handling extracted data.
Capability inventory: The skill includes scripts that can perform file writes to the local system (e.g., via the --output flag) and documentation demonstrating the use of subprocess.run to chain operations.
Sanitization: While the documentation provides examples for validating structured data like forms (using jsonschema), there is no evidence of broad sanitization for unstructured text extracted from PDFs.
[COMMAND_EXECUTION]: The skill documentation and workflows rely on executing shell commands and Python scripts via subprocess.run.
Evidence: SKILL.md includes a batch processing workflow using subprocess.run to call internal scripts. OCR.md provides instructions for installing system dependencies using sudo apt-get install and brew install. While these are legitimate setup and operational steps, they involve the execution of system-level commands.

PDF Processing Pro