The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently generates shell commands by interpolating user-controlled variables such as <pdf>, <epub>, and <workdir> into command templates for tools like pdfinfo, pdftk, mutool, pdftotext, pandoc, and ocrmypdf. The assembly of these commands through simple string concatenation without rigorous sanitization or escaping of file paths allows for command injection if a user or a malicious file system provides a path containing shell metacharacters (e.g., ;, |, `).
[PRIVILEGE_ESCALATION]: Documentation in references/output-modes.md explicitly instructs the user to execute administrative commands using sudo (e.g., sudo tlmgr update --self). Providing instructions for privilege escalation within the skill's reference material increases the risk of unauthorized system-level changes if an agent attempts to automate these setup steps.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes large-scale untrusted text from books and public URLs, creating a surface for indirect prompt injection. A malicious book could contain hidden instructions designed to manipulate the agent's summarization or subsequent actions.
Ingestion points: File paths for PDF, EPUB, MOBI, Markdown, and TXT files; URLs for public-domain works (SKILL.md, references/sources.md).
Boundary markers: None identified; the skill does not use specific delimiters or instructions to ignore embedded commands in source material.
Capability inventory: Shell command execution (pandoc, pdfinfo, etc.), file system write access (~/Documents/books/), and network fetching (WebFetch).
Sanitization: None identified for the ingested book content.
[EXTERNAL_DOWNLOADS]: The skill facilitates downloading book content from external services including Project Gutenberg and Archive.org via the WebFetch tool. These are well-known services for public-domain literature, and the activity is consistent with the skill's primary purpose.
[DYNAMIC_EXECUTION]: The skill utilizes the --render flag to dynamically generate PDF or HTML documents from generated markdown notes using pandoc. This process involves executing external formatting engines on content the agent has compiled, which requires execution privileges over local tooling.

read-book