The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted book files (EPUB and PDF) which could theoretically contain malicious instructions designed to influence the AI's behavior.
Ingestion points: path/to/book.epub and path/to/book.pdf are used as inputs for text extraction in SKILL.md.
Boundary markers: The skill architecture uses a "Trust contract" where child subagents handling untrusted content are restricted to read-only tools (get_page, search).
Capability inventory: The orchestrator can write results to the repository via put_page and execute shell/Python commands.
Sanitization: The skill performs basic text extraction and HTML tag stripping to clean content before analysis.
[DYNAMIC_EXECUTION]: The skill uses shell scripts to dynamically generate and run Python code for processing book content.
Evidence: A Python script is defined as a heredoc and executed via `python3
<<'PY'in the EPUB extraction section ofSKILL.md`.
[EXTERNAL_DOWNLOADS]: The skill references external software packages required for operation.
Evidence: Instructs the installation of beautifulsoup4 and lxml via pip3. These are standard, well-known libraries for document parsing.
[COMMAND_EXECUTION]: The skill relies on command-line utilities for file manipulation and conversion.
Evidence: Uses unzip, find, mktemp, and pdftotext to extract and prepare book text.

book-mirror