The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data (images and PDF files), which creates a surface for indirect prompt injection. If an image or PDF contains malicious text, the agent might interpret the extracted content as instructions rather than data. Additionally, the 'Post-Execution Reflection' instructs the agent to modify its own skill files based on its findings, which could lead to accidental or malicious instruction corruption if the OCR output is flawed.
Ingestion points: The --input parameter in the unlimited_ocr.py script accepts untrusted binary files including images and PDFs.
Boundary markers: The skill instructions do not define delimiters or specific 'ignore' directives to separate the extracted OCR text from the agent's operational instructions.
Capability inventory: The skill has permission to use Bash, Read, Write, and Glob tools, allowing it to modify files and execute system commands.
Sanitization: There is no documented process for sanitizing or escaping the text extracted from documents before it is integrated into the agent's context.
[COMMAND_EXECUTION]: The skill executes a local Python script (unlimited_ocr.py) using the uv run utility. This execution involves handling file paths and directory structures provided as user input.
[EXTERNAL_DOWNLOADS]: Fetches machine learning model weights and configuration files from HuggingFace repositories, specifically baidu/Unlimited-OCR and mlx-community/Unlimited-OCR-mxfp8.

unlimited-ocr-parse-document