The Agent Skills Directory

[SAFE]: The skill performs legitimate image-to-text processing for tax-related documents. It includes logic for handling PDF text extraction and fallback to image processing when necessary.
[COMMAND_EXECUTION]: Uses internal utility commands (shinkoku pdf extract-text and shinkoku pdf to-image) which appear to be part of a local domain-specific toolset. These are used for their intended purpose of document parsing.
[INDIRECT_PROMPT_INJECTION]: While the skill processes untrusted external data (images/PDFs provided by users), it implements a 'dual-verification' strategy where two independent AI contexts compare results. This significantly mitigates the risk of the model following instructions embedded in the documents (OCR-based injection) by requiring consistency between two separate runs or manual user confirmation upon discrepancy.

reading-deduction-cert