The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute several Python scripts (scripts/ocr_caller.py, scripts/configure.py, scripts/smoke_test.py) to perform OCR tasks, manage configuration, and test connectivity. These scripts take user-provided arguments like file paths and URLs directly into command-line parameters.
[EXTERNAL_DOWNLOADS]: The skill is designed to send local files or fetch content from external URLs to be processed by a remote OCR service defined by the API_URL environment variable. This involves outbound network communication with a third-party provider.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources (images, PDFs, and URLs) and returns the extracted text to the agent's context. There are no explicit boundary markers or sanitization steps mentioned in the documentation to prevent the LLM from obeying instructions hidden within the OCR results.
Ingestion points: User-provided inputs via --file-url and --file-path parameters in SKILL.md.
Boundary markers: None identified in the provided documentation.
Capability inventory: Execution of Python scripts, network requests to external APIs, and local file reading.
Sanitization: Not present; the raw extracted text is returned as result.full_text.
[DATA_EXPOSURE]: The skill provides a mechanism to read local filesystem paths (--file-path) and transmit data to an external API endpoint. It also handles sensitive information like PADDLE_OCR_TOKEN, which is stored in a .env file.

ppocrv5