The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests and processes untrusted external data in the form of academic abstracts and conclusion excerpts fetched from Semantic Scholar and arXiv. This data is interpolated into prompt templates (such as references/classify.md and references/detail.md), creating a surface for indirect prompt injection if malicious instructions are embedded within paper text.
Ingestion points: Data enters via scripts/web_api.py and scripts/paper_md.py.
Boundary markers: The templates utilize numbered lists and section headers to organize input, but they do not include explicit instructions to the model to ignore embedded directives in the paper content.
Capability inventory: The skill requires execute and write_file permissions, which are used to run its internal CLI and save the final report.
Sanitization: The skill applies character escaping to ensure Mermaid diagram integrity in scripts/mermaid.py but does not perform security-focused sanitization on the natural language data.
[EXTERNAL_DOWNLOADS]: The skill initiates network requests to reputable academic sources including api.semanticscholar.org, arxiv.org, and content conversion services like arxiv2md.org and markxiv.org to retrieve research metadata and full-text excerpts.
[COMMAND_EXECUTION]: The runbook instructs the host agent to execute its own Python CLI dispatcher (scripts/cli.py) using uv run to perform deterministic parsing, formatting, and rendering tasks.

paper-graph