The Agent Skills Directory

[SAFE]: The skill represents a benchmark for secure engineering practices. It provides thorough guidance on secret management (using Vault or cloud providers), data isolation, and defensive coding.
[INDIRECT_PROMPT_INJECTION]: The skill defines a significant attack surface as its primary purpose is ingesting untrusted data (PDFs, images, Excel files, and external API responses).
Ingestion points: External file uploads (PDF, Image, Excel) and third-party API payloads (referenced in SKILL.md and references/pdf-extraction.md).
Boundary markers: The skill mandates strict validation boundaries using Pydantic models (references/validation-and-deadletter.md).
Capability inventory: The framework utilizes subprocess.run for system tools (OCR, scanning), sqlalchemy for database operations, and httpx/stripe for network communication.
Sanitization: Proactive sanitization is implemented, including EXIF stripping (piexif/Pillow), PDF sanitization (qpdf), and strict MIME-type validation (python-magic).
[COMMAND_EXECUTION]: System tools like ocrmypdf, pdftoppm, and clamdscan are invoked via subprocess.run in references/pdf-extraction.md. These are implemented using list-based arguments for data processing and security scanning, adhering to safe execution patterns.

python-data-pipelines