The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses curl to interact with academic APIs and echo to append JSON-formatted telemetry to a local execution log file (~/.claude/PAI/MEMORY/SKILLS/execution.jsonl).
[EXTERNAL_DOWNLOADS]: It fetches paper metadata and AI-generated overviews from export.arxiv.org and alphaxiv.org. These are well-known services within the academic research community.
[DATA_EXFILTRATION]: Usage metadata is stored in a local directory for persistence. No sensitive user data or credentials are transmitted to external servers.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the content of academic papers (abstracts or AlphaXiv summaries) it processes.
Ingestion points: Atom XML from arxiv.org and Markdown overviews from alphaxiv.org in all workflow files.
Boundary markers: Absent; fetched content is presented directly to the agent without isolation delimiters.
Capability inventory: The agent can execute curl for network requests and write to the local file system via echo redirection.
Sanitization: Input query terms are URL-encoded, but the fetched paper content is not sanitized for potential embedded instructions.

ArXiv