The Agent Skills Directory

[SAFE]: The skill implements a rigorous source discipline, requiring the agent to explicitly distinguish between original paper text, related literature, and inferences. This core design promotes high-fidelity summarization and reduces hallucination.
[PROMPT_INJECTION]: The skill ingests data from external papers and URLs, creating a surface for potential indirect prompt injection. Ingestion points: PDFs, arXiv IDs, and web URLs. Boundary markers: The instructions mandate clear labeling of source content using tags like 【论文原文】 and separate inference from fact. Capability inventory: The skill can read local files and generate self-contained HTML/JavaScript files. Sanitization: The agent is directed to follow an evidence-led study workflow which serves as a reasoning-based filter for external content.
[SAFE]: The skill generates interactive HTML/JavaScript code to provide 'Explanatory UI' visualizations. This is a standard capability for high-end agent skills to explain complex concepts (e.g., parameter effects, geometry). The instructions specify that this code must be self-contained with no external dependencies and provide static fallbacks (SVG or tables) when interactive environments are unavailable.

summarize-paper