The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the docling CLI and the uv tool to perform document conversions, install dependencies, and run a local evaluation script (scripts/docling-evaluate.py). These operations are within the expected functional scope of a document processing utility.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest and process external documents (PDF, HTML, images) from local paths or URLs. This represents an indirect prompt injection surface where malicious content within a document could attempt to influence the agent's behavior during subsequent processing steps.
Ingestion points: docling processes user-supplied files and remote URLs (e.g., Arxiv links).
Boundary markers: The instructions do not specify explicit delimiters or warnings for the agent to ignore instructions embedded within the converted text.
Capability inventory: The skill has the ability to write to the file system (--output), execute local CLI commands, and perform network requests for remote VLM processing.
Sanitization: While Docling performs internal normalization of document structures, there is no specific sanitization mentioned to protect against instructions intended for the LLM.
[CREDENTIALS_UNSAFE]: The skill documentation includes examples for configuring remote VLM APIs that require authorization headers. However, it correctly uses placeholders like YOUR_KEY instead of hardcoding any actual secrets, which is considered safe practice.

docling-document-intelligence