The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill uses curl to download PDFs, archives, and HTML content from arxiv.org. While ArXiv is a trusted and well-known academic repository, fetching external content creates a data ingestion surface that the agent must process.
[COMMAND_EXECUTION]: Local Python scripts (generate_note.py and update_graph.py) are executed via shell commands defined in SKILL.md. These commands interpolate user-controlled or paper-controlled variables like $TITLE and $AUTHORS directly into the shell execution context.
[PROMPT_INJECTION]: Indirect Prompt Injection Surface. Metadata extracted from untrusted external papers flows into shell arguments and Markdown templates. A paper with a malicious title or author field containing shell control characters or Markdown injection sequences could attempt to influence the agent's behavior or corrupt the formatting of the generated research notes.
Ingestion points: ArXiv paper content and metadata retrieved via network requests in SKILL.md.
Boundary markers: None identified. Metadata is used directly in shell command templates and f-string file generation.
Capability inventory: Shell command execution, local file system writes, and modification of structured JSON data.
Sanitization: generate_note.py includes basic character replacement for filenames, but it does not sanitize metadata used inside the Markdown content or within the shell command invocation defined in SKILL.md.

paper-analyzer