The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It extracts text from untrusted external sources (images and PDFs) and interpolates that text directly into prompts for an LLM (Claude) to generate summaries, answer questions, and detect document types.
Ingestion points: The input_path parameter in scripts/doc_pipeline.py, scripts/structured_extract.py, and scripts/table_extract.py serves as the entry point for external data.
Boundary markers: While the prompts use headers like 'DOCUMENT:' and 'TEXT:', they lack robust delimiters or specific instructions to the LLM to ignore embedded commands, increasing the risk of the model following instructions contained within the analyzed documents.
Capability inventory: The skill has the capability to perform network requests (to the Anthropic API) and write results to the local file system.
Sanitization: No sanitization, filtering, or escaping is performed on the extracted OCR text before it is inserted into the LLM prompts.
[EXTERNAL_DOWNLOADS]: The skill downloads official pre-trained models from PaddleOCR's infrastructure to a local cache (~/.paddlex/official_models/). It also performs network operations to the Anthropic API for document reasoning. These are standard operations for the skill's intended functionality and target well-known, trusted services.

paddleocr