The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skills ingest academic metadata from external sources like CrossRef, OpenAlex, and arXiv. This metadata is untrusted and could contain malicious instructions designed to influence an agent processing the references.\n
Ingestion points: Multiple skills (e.g., academic-citation-manager/SKILL.md, bibtex-management-guide/SKILL.md) use functions like get_bibtex_from_doi and parse_bibtex_entries to fetch and parse external strings from academic APIs.\n
Boundary markers: The provided code templates do not include explicit prompt delimiters or instructions to ignore instructions embedded in the processed bibliographic data.\n
Capability inventory: The skills possess capabilities for network requests (requests.get) and file management via the included Python code templates.\n
Sanitization: While basic character replacement and regex-based cleaning are used for formatting, no specific protection against indirect prompt injection is implemented.\n- [EXTERNAL_DOWNLOADS]: The skill guides provide instructions for installing external libraries and tools such as onecite, papis, and various Zotero plugins. These instructions point to official registries (PyPI, NPM) or established GitHub repositories (e.g., JabRef/jabref, papis/papis).\n- [COMMAND_EXECUTION]: The skills provide Python scripts and terminal command examples for interacting with academic APIs and managing local bibliography files. These capabilities, while necessary for the stated purpose, represent an interaction surface for the agent with potentially untrusted external data.

citation-skills