The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and parse untrusted data from various document formats (PDF, DOCX, XLSX, etc.) which are then processed by the agent. \n
Ingestion points: The LiteParse.parse method and the lit parse CLI command in SKILL.md and scripts/batch_parse_dir.py ingest data from local files or standard input.\n
Boundary markers: The extracted text is returned without specific delimiters or warnings to the agent to ignore instructions embedded within the document content.\n
Capability inventory: The skill environment permits file system operations and shell command execution through the allowed tools, providing a surface for impact if an injection is successful.\n
Sanitization: The skill does not perform sanitization or filtering of the extracted text to prevent malicious instructions from reaching the agent context.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of external packages and system-level tools from established sources.\n
The instructions specify installing liteparse from PyPI and @llamaindex/liteparse from npm, which are official packages from LlamaIndex.\n
It also guides the user to install system dependencies like LibreOffice and ImageMagick for extended format support using standard package managers.\n- [COMMAND_EXECUTION]: The skill provides a CLI tool and a helper script for batch processing documents.\n
SKILL.md and references/cli_reference.md demonstrate the use of the lit CLI tool for parsing and screenshots.\n
scripts/batch_parse_dir.py is a Python utility that performs file system operations and calls the parsing library to process directories of documents.\n- [DATA_EXFILTRATION]: The skill supports sending document data to an external server for OCR processing.\n
The ocr_server_url parameter documented in references/ocr_and_formats.md and references/api_reference.md allows document images to be sent to an external HTTP endpoint. Users should ensure any configured OCR server is trusted.

liteparse