The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a set of local bash scripts to coordinate the reading workflow. It relies on standard document processing utilities such as pdftotext (via Poppler) and pandoc to convert user-supplied document formats into a paged plain-text format for the agent to process. These operations are local and confined to the user-provided files and a generated workspace directory.
[INDIRECT_PROMPT_INJECTION]: By processing long-form external documents like PDFs and EPUBs, the skill interacts with untrusted data that could theoretically contain prompt injection attacks. The skill mitigates this risk through its multi-step architecture: it breaks documents into manageable segments, records observations in intermediate note files, and provides a verify-quotes.sh utility to mechanically validate the agent's syntheses against the original source text.
[DATA_EXPOSURE]: The skill creates and maintains a local directory (<slug>-notes/) to store the coordinate-mapped source text and chapter-level notes. It does not access sensitive system paths, hardcode credentials, or perform any network operations, ensuring that the processed data remains within the local environment.

deep-reader