The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several local bash scripts (prepare-text.sh, split-chapters.sh, read-pages.sh, etc.) to process documents. These scripts are invoked via shell commands to convert formats (PDF/EPUB to text), manage a local workspace, and slice the content into manageable pieces for the AI agent.
[DATA_EXPOSURE]: The skill maintains all processed data within a local workspace directory (e.g., <slug>-notes/) created next to the source file. It does not perform any network operations or attempts to access sensitive system paths outside of the specified workspace.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied external files (PDFs, EPUBs, etc.) which is an attack surface. However, it mitigates this risk by converting content to plain text, using explicit boundary markers ([[page N]]), and processing the document in discrete passes rather than loading the entire untrusted content into the context at once. The "Recite" and "Verify" steps further instruct the agent to check propositions against the source text, providing manual and mechanical guardrails.
Ingestion points: User-supplied files (PDF, EPUB, DOCX, TXT, MD) are processed by prepare-text.sh.
Boundary markers: The skill inserts page anchors in the form of [[page N]] and splits chapters into standalone files in chapters/ to keep content segmented.
Capability inventory: The skill uses bash, awk, pdftotext, and pandoc for file processing and workspace management.
Sanitization: Input filenames are slugified (sanitized of special characters) before directory creation, and content is normalized to plain text via established conversion tools.
[EXTERNAL_DOWNLOADS]: The documentation mentions using npx to install the skill from a GitHub repository (tronghieu/agent-skills). This is the standard mechanism for the platform and originates from the identified author.

deep-reader