The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: Indirect Prompt Injection Surface
Ingestion points: Ingestion occurs via the load_structure function in scripts/_common.py, which reads local external materials files (e.g., CIF, POSCAR) provided as inputs to various analysis and conversion scripts.
Boundary markers: No explicit prompt boundary delimiters or protective instructions wrap the final serialized JSON text when it is output to the agent's context window.
Capability inventory: The skill environment exposes powerful tools including Read, Write, Bash, Glob, and Python code execution.
Sanitization: Although the scripts enforce robust structural and numerical bounds (such as maximum site counts, coordinate types, and file sizes), text descriptions, labels, and parser warning messages from untrusted files are echoed without string sanitization.
[EXTERNAL_DOWNLOADS]: Safe External Service Connections
Description: The skill connects to the official Materials Project API (api.materialsproject.org) to fetch materials data via the established mp-api client when explicitly authorized by the user. It also makes references to fetching citation records from the arXiv academic repository (arxiv.org). These interactions involve well-known scientific repositories and employ secure access patterns, including environment-variable token loading and explicit credential redaction.

pymatgen