The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes scientific PDF papers which are untrusted inputs. Maliciously crafted content within these documents could potentially influence the LLM's behavior during the extraction process.
Ingestion points: The scripts/extractor.py script reads the content of local PDF files using either the PyMuPDF library or the Mathpix API.
Boundary markers: In scripts/extractor.py, the extracted content is interpolated into the LLM prompt using triple-quote delimiters ('''{\"\"\" {content} \"\"\"}'''). This provides a basic level of separation but may be bypassable by content containing specifically crafted escape sequences.
Capability inventory: The skill has access to Bash, Read, and Write tools; it writes extracted data to the local filesystem using the OutputHandler.save_output method.
Sanitization: While the script includes a remove_references method to strip bibliography sections, it does not perform sanitization to detect or filter potential prompt injection instructions within the document text.
[EXTERNAL_DOWNLOADS]: The skill communicates with external APIs to perform its primary functions.
Evidence: scripts/extractor.py sends PDF data to api.mathpix.com for OCR and utilizes the openai Python client to send extracted text to LLM endpoints (defaulting to Anthropic's API). These are well-known technology services required for the skill's stated purpose.
[COMMAND_EXECUTION]: The skill provides Python scripts designed to be executed via the Bash tool to automate scientific research workflows.
Evidence: The skill includes scripts/extractor.py and scripts/batch_extract.py which are invoked by the agent to process files and save results to the filesystem.

sci-data-extractor