The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/pdf_ocr_processor.py uses subprocess.check_call to execute shell commands for package installation. While the package names are hardcoded, this allows the skill to modify the system's Python environment at runtime.
[EXTERNAL_DOWNLOADS]: The skill is designed to dynamically download and install Python packages (such as rapidocr_onnxruntime, paddleocr, and rapid-doc) from the public PyPI registry if they are not detected on the system during execution.
[DATA_EXFILTRATION]: When using the SiliconFlow OCR engine, the skill transmits document image data to an external cloud service (api.siliconflow.cn). This is the intended functionality for cloud-based recognition but involves sending potentially sensitive file content to a third-party API.
[PROMPT_INJECTION]: The skill ingests untrusted text data extracted from PDFs and images. Because this data is returned to the agent's context without sanitization or boundary delimiters, it creates a risk of indirect prompt injection where instructions embedded in a scanned document could hijack the agent's behavior.
Ingestion points: scripts/pdf_ocr_processor.py (via ocr_pdf and ocr_image_file methods) reads content from PDF and image files provided to the agent.
Boundary markers: The SiliconFlowOCREngine uses a prompt to request text output, but the resulting content is merged into the agent's context without specific "ignore instructions" delimiters.
Capability inventory: The script performs file reads, network POST requests, and subprocess shell execution.
Sanitization: While the RapidDocEngine performs basic regex cleaning to remove markdown formatting, no generic sanitization is applied to the raw text output across all engines.

pdf-ocr-skill